Other games lack such a salient point of agreement, though, and for them there is no argument as decisive as there is for
Table 9.3 Outcome <8, 2> <1, 10> <5, 4> Concession for Paul 8 − 8 = 0 8 − 1 = 7 8 − 5 = 3 Concession for Lucy10 − 2 = 810 − 10 = 010 − 4 = 6
Table 9.4 Outcome <8, 2> <1, 10><5, 4> Relative Concession for Paul 0/8 = 07/8 3/8 Relative Concession for Lucy8/9 0/9 = 0 6/9 = 2/3 Gauthier claims that the rational agreement is the one that minimizes the maximum concession. Find the outcome whose maximum relative concession is smallest. Call this point the Minimax Relative Concession point, or MRC. 14 That is the rational point of agreement. In our example, <5, 4> is the MRC, since the maximum relative concession in its column is Lucy's concession of 2 3 , and the other columns have higher maximum relative concessions. So, according to Gauthier, it is the solution that it would be rational for Lucy and Paul to agree on.
This is not obvious. 15 But Gauthier has an argument for it. 16 If some point other than the MRC point is chosen, one of the parties may complain that she is being asked to concede too much. The MRC point has this feature: no alternative would require less relative sacrifice by the person asked to sacrifice most, where sacrifice is measured by relative concession.
The Power of Gauthier's Theory When the story about rational bindingness is combined with the story about how to find a rational agreement, the result is very powerful. It explains why we should cooperate, why it is irrational not to cooperate, and it explains what cooperation involves. If successful, the explanation, although it is not apt to explain our intuitive commonsense morality, will manage to explain certain compelling features of commonsense morality, and, even more important, it will explain why and in what sense we have reasons to comply with these moral demands. Gauthier's theory is also important, to my mind, for the criticism, discussion, and commentary that it has inspired. I conclude this section by briefly surveying some of the criticisms.
Is Constrained Maximization Possible? If we understand utility and preference in the standard way of decision theory, it is very hard to see how it is even possible for someone to avoid being a straightforward utility maximizer. Decision theory constructs a utility function for each agent in such a way that (a) a given option will have a higher utility than a second if and only if the agent prefers it to the second, and (b) the utility for each option will be the same as its expected utility. Imagine that Connie is utterly convinced by Gauthier's argument and decides to become a constrained maximizer. She is no longer a straightforward maximizer. So on some occasion she chooses the option with the lower expected utility. But by definition, the option with the higher utility is the one Connie prefers! So, a decision theorist will say instead that the utility function used to represent Connie's preferences was the wrong one. Since she just chose (for instance) to Cooperate with her partner in what appeared to be PD, her actual utility function must be one that assigns a higher utility to Cooperating than it assigns to Defecting. As long as Connie's preferences satisfy the axioms of decision theory, there is a new function that represents her preferences by an expectational utility function. So she is still a straightforward maximizer. Now, presumably something happened when Connie decided (as she thought) to become a constrained maximizer. Her behavior changed; her disposition to cooperate with others changed. If we cannot describe the change by saying that Connie stopped being a straightforward maximizer, how can we understand it instead? Roughly speaking, Connie acquired a tendency to cooperate with others even when cooperation would not maximize the expectation of her old utility function. She now prefers, it seems, to cooperate, even when defecting has a higher utility according to her old preferences. 17
Connie has changed utility functions by changing preferences, as people often do. As a result, the game she is now playing is not PD at all. PD no longer models correctly her interaction with a partner. 18 end p.167
Restricting the Set of Preferences What might be thought to be most interesting about Gauthier's argument is that it shows how someone might start from purely selfish preferences and become convinced that she should develop the new dispositions (the dispositions to cooperate with other conditional cooperators). It would be preferable, from her own selfish point of view, to have the nonselfish tendencies. I do think that this is an interesting feature of the argument. But it is not completely obvious why it is interesting. After all, most people are not utterly selfish. Is it of merely theoretic interest that a perfectly selfish person could be convinced (under the right circ*mstances) to change her stripes? Or is there some practical import, too? Suppose we construct for Connie (whom we will not presume to be utterly selfish) a special utility function, one that represents not her full preference structure but only her selfish preferences. Let's call this her s-utility function. Now we can think of Gauthier's argument as showing that Connie's s-utility will be higher in the long run if she acts as a constrained s-utility maximizer. 19 Why is this interesting? Well, it may show something about some of Connie's other, nonselfish preferences. It may show that some of her preferences to cooperate actually contribute in the long run to the satisfaction of her selfish preferences. In that case, even in her more selfish moments she can reflectively endorse those nonselfish preferences. 20 It is an important bit of moral psychology to show that certain of our “ethical” preferences are supported and endorsed from the point of view of our more selfish preferences.
The Baseline Let's look at the question of exactly what is to count as a move in the game. For instance, in the story about Lucy and Paul, the various ways they could work together were counted as possible moves, and so was working separately. But we did not consider the “move” of Lucy's murdering Paul and taking all of his money, or enslaving him at gunpoint and forcing him to work for her. That these were not counted as moves was a critical fact in the subsequent determination of the MRC point of the game, since relative concession is measured by using the yardstick of the difference between favorite outcome and fallback position (no agreement). Why do we count working separately as the fallback position, rather than enslaving the other party, or perhaps being enslaved by the other party?
Gauthier rules out things like enslaving or killing other parties by a version of Locke's famous Proviso. 21 Gauthier's version “prohibits worsening the situation of another person, except to avoid worsening one's own through interaction with end p.168 that person.This, we claim, expresses the underlying idea of not taking advantage” (Gauthier 1988 , 205). Enslaving another worsens his situation, of course, as does killing him. But we need more precision. Polluting (an example much discussed by Gauthier) also worsens the situation of others. Is polluting ruled out as a possible move in any interactive game? No. Polluting does worsen the situation of another, but not “through interaction,” as Gauthier understands it. Don Hubin and Mark Lambeth (1991 , esp. 114– 16) give a more precise explanation. The Proviso comes into play whenever I do something that makes you worse off than you would have been had I not existed. It does not, however, debar me from all such actions. Suppose that if I don't do it, then I will be worse off than I would have been had you never existed. Then the Proviso no longer applies. Understood this way, the Proviso does express the idea that I must not exploit you. 22 The Proviso allows me to pollute. Gauthier gives this example: if I live upstream from you, I may reasonably throw my wastes into the stream even though you suffer for it, because had you not been downstream from me I would have polluted anyway. On the other hand, I may not start poisoning the river for the sole purpose of getting you to pay me to stop. That would be taking advantage of you (1988 , 211). There are some serious questions about whether Gauthier's Proviso captures our intuitive idea of natural rights, but for Gauthier's own purposes, that may not matter much. What does matter is that agreements reached from the Proviso's baseline are rational to keep, while agreements reached by negotiating from “inadmissible” starting points are not. Gauthier's argument is that it is not rational to be disposed to comply with predatory, exploitive bargainers. I will be better off, in the long run and given that others are rational, if I am disposed to refuse to comply with or even bargain with exploitive predators. 23
Concluding Thoughts on Gauthier Gauthier's grand project of constructing morality from (self-)interests has produced so much interesting and fruitful commentary that even if it turns out to be an utter failure it will have been worthwhile. 24 The apparatus of decision theory features prominently as a tool in the construction. For the reasons given I doubt that some of Gauthier's main conclusions make sense as long as the concepts of decision theory are interpreted in their standard way. Some adjustments of interpretation can help to preserve the coherence of the theory. In the next sections I consider more significant departures from the standard interpretation of decision theory. What happens if we try to understand decision theory as being about what is good and bad, rather than about what people prefer?
end p.169
3. The Good It can be useful to think of decision theory not as a theory of what it is rational to do, but rather as a formal theory. By “formal theory” I mean one left uninterpreted, so that we have just a set of axioms, and we can draw out consequences of those axioms by purely formal means. If we can then find interpretations under which the axioms turn out to be plausible, we will then have a reason to believe the theorems, the consequences of the theory. It is, as I have said, a mistake to think of decision theory proper as a theory of the good for a person. Be that as it may, decision theory viewed as a formal theory may still be quite relevant to straightforward ethical questions about what is good for a person, and also to questions about what is morally good (and right). We may ask whether the axioms of the theory are plausible if they are interpreted as being about which alternatives are better than which for one or another person, or about which alternatives are morally better than which. In the next section I will present a version of Harsanyi's theorem. In the version I'll present, adapted from Broome 1991 , it purports to be an argument from apparently weak premises to the conclusion that (something very like) utilitarianism is true. I will conclude by explaining which of the premises should seem most questionable to those who find utilitarianism unattractive.
Good for Persons We want to know whether the relation, “at least as good for a person as,” satisfies the constraints given in the axioms of decision theory. That is, if we interpret “xRy” as meaning “x is at least as good for this person as y is,” do the axioms come out true? We can start with the ordering axioms. One might raise questions about completeness. It is at least arguable that becoming an astronaut will not be better for you than becoming a concert violinist, nor will it be worse, nor exactly as good. While “at least as good as” is not obviously complete, I will proceed as if it were. There are degrees of completeness. It may be that a very large field of important prospects can be ordered under “at least as good as,” in which case what follows will apply to goodness for persons restricted to that field. Since reflexivity is trivially true of “at least as good as,” we can turn to transitivity. It has occasionally been argued that the “better than” relation (for persons) is not transitive. 25 But it must be transitive, for “better than” is a comparative. It means “has more good than.” And “more” is logically transitive. If intuitions seem to point to the failure of transitivity of “better than,” then either (a) the intuitions are misleading, or (b) the intuitive relation has been misnamed. end p.170
The lottery axioms are trickier. Are they true, in our interpretation? The reduction of compound lotteries may seem doubtful simply on the grounds that in some circ*mstances, having to sit through a lottery could itself be a bad thing. Suppose you could either have a simple lottery offering you a 1 2 chance at a new philosophy book and otherwise a tuna sandwich, or else a hugely complex lottery with hundreds of sublotteries nested inside, each as prizes in the lotteries in the levels above, but with the same ultimate prizes (and same chances for them). It's better to have the simple lottery, unless you happen to enjoy the gambling experience and have plenty of free time. We can finesse this problem. For the purposes of formal decision theory, it is not necessary to think of the lotteries as familiar gambling events. They go on “behind the scenes,” as it were; you do not experience them at all as events in their own right. So you won't notice the difference between the compound lottery and its formally equivalent simple lottery. Continuity presents a more serious issue. Suppose that at the moment, Jessie would enjoy eating a banana. Compare the following prospects: B She gets the banana. C She continues as she is now, without the banana. D She dies immediately. Plausibly those prospects are properly ranked from better to worse for Jessie, B better than C better than D. Now continuity says that there is some lottery between B and D that is exactly as good for Jessie as C is: call it “L[p, B, D].” What probability is p? Many people think that there is no such p (that is, no such p < 1; when p = 1, the “lottery” is strictly better for Jessie). It is somewhat plausible, at least, that there are some goods that are lexically prior to others, in the sense that it is worse to take any chance of losing the higher-ordered good in exchange for any chance of gaining the lower-ordered. This may be a real problem. But even if it is real, it may be limited. The “at least as good as” relation might satisfy the axioms over a limited domain, and a separate one might satisfy them again over another, “lexically higher” domain. If there are many different orders of goods, then the application of decision theory to the good for a person will be of relatively little use. But if there are only two, say, then the rich structure of decision theory will still be preserved within the domains. There will be a lower-order expectational utility function representing how good things are for a person, and a higherorder one. Utility space would be two-dimensional. 26 That would be interesting, not disastrous for the model. I am unaware of any particular difficulties in supposing that “at least as good as” satisfies the better chances axiom. There is a well-known objection to the better prizes axiom, due to Maurice Allais (see Sorensen, chap. 14, this volume). 27 However, this objection is complicated to present. Since I am going to discuss a closely related problem later, when we consider whether the relation “morally at least as good as” satisfies this same axiom, I will put off consideration of better prizes until then.
If the axioms are true under the personal good interpretation, then so must be whatever theorems follow. In particular, the Expected Utility Theorem must be. That is, if the axioms are true, then for each person there will be an expectational utility function that represents that person's good. Actually, there will be a whole family of such functions, each a positive linear transformation of the others. The representability of a person's good by an expectational utility function is interesting for a couple of reasons. First, it provides a measure of goodness for persons. We can now ask not only whether one option is better for you than a second, but also how much better. Once a standard unit is fixed, that question will have an answer. It would make sense to say that the change in a person's well-being was twice as great this year as it was last year; that ratio, since it is a ratio of differences, would be independent of the choice of scale (just as the ratio of differences in temperature remains constant no matter which scale we use). While this result is interesting, more significant is the role it plays in the argument of the next section. Together with some similar assumptions about moral goodness, the assumption that the “at least as good as” relation for persons satisfies the axioms of decision theory turns out to imply a strikingly strong conclusion.
Moral Goodness Things can be better or worse for persons. More controversially, they can also be simply good or bad. Some philosophers doubt that there is any such thing as a “good state of affairs,” 28 but we do seem to have some conception of things turning out better or worse from the moral point of view. In preparation for the main argument, I will ask first whether the relation “at least as morally good as” satisfies the axioms of decision theory, and second whether an option must be (morally) at least as good as another if it is at least as good for everyone. If the answer to both is yes, then a third condition will imply a very strong conclusion, namely a form of utilitarianism. Let's call the “at least as morally good as” relation “M”. Now there is a very quick argument to the conclusion that M does satisfy the axioms. Imagine someone, call her Angela, who is concerned with nothing other than the moral state of the world. For every pair of alternatives, including lotteries, Angela always prefers the one that is morally better and is indifferent between those that are equally good. Then the ordering imposed by Angela's preference relation is also the ordering of states of the world by M. Unless Angela is irrational by decision theoretic standards, her preference relation satisfies the axioms; then so does M. end p.172 There is, however, a particular doubt about the better prizes axiom that arises specifically in moral contexts. Suppose that option 1 is definitely better than option 2 in some case, and at least as good in every case. Then the axiom says that option 1 is strictly better than option 2. Dominance principles, like better prizes, seem very plausible. How could option 2 be as good as option 1 if it could not turn out better and might very well turn out worse? It is hard to see, in the abstract, how something like that might happen. But when the
relation in question is a moral relation, there is a special reason to be wary. This reason has to do with fairness, a specifically moral property. Sometimes an option can be fair exactly because it might turn out one way and might turn out the other way. The fairness resides in the openness of the possibilities, and not locally in one possible outcome or the other. Considerations of fairness jeopardize better prizes. Here is an example. 29 Suppose you have a kitten, which you plan to give away to either Talia or Horace. Talia and Horace both want the kitten very much. Both are deserving, and both would take good care of the kitten. You are sure that giving the kitten to Talia is at least as good as giving it to Horace. But you think that would be unfair to Horace. You decide to flip a fair coin: if the coin lands heads, you will give the kitten to Horace, and if it lands tails, you will give the kitten to Talia. We can call this the fair chance option. Let's use h for “Horace gets the kitten” and t for “Talia gets the kitten.” Since giving the kitten to Talia is at least as good as giving it to Horace, (K1) tRh Better prizes tells us (L2) (x)(y)(z)(p)(xRy ↔ L[p, z, x] R L [p, z, y]) so it implies this existential instantiation (K2) (z)(tRh ↔ L[1 2 , z, t] R L[1 2 , z, h]) Now (K1) and (K2) together imply (K3) (z)L[1 2 , z, t] R L[1 2 , z, h] But now instantiate the variable z with t: (K4) L[1 2 , t, t] R L[1 2 , t, h] The first option, L[1 2 , t, t], is the option of flipping the coin and giving the kitten to Talia no matter which side turns up. The second option is the fair chance option. Giving the kitten to Talia is not as good as the fair chance option, so (K4) is false. So better prizes is false, under the present interpretation.
The kitten story seems to be an example of how dominance can fail. We can make this clearer by supposing that giving the kitten (outright) to Talia is just end p.173 slightly better than giving it to Horace. Maybe Talia will be able to take somewhat better care of the kitten. It is still plausible that the fair chance option is morally better than giving the kitten outright to Talia. But giving the kitten outright to Talia dominates the fair chance option, since it is strictly better in some case and at least as good in every case. This is a problem for better prizes. If it is insurmountable, then moral goodness cannot be represented by an expectational utility function. Before concluding the moral goodness is not expectational, we should take a closer look at the kitten story. The fair chances option is better than giving the kitten to Talia because it is fair. Fairness, I said, does not reside in the outcome but in the chanciness of the lottery itself. Dominance principles seem plausible when it seems reasonable that all of the goodness in a complex must reside in the simple parts of that complex. The fairness of lotteries, though, does not reside in the outcomes of the lotteries. Or does it? I am not so sure. Compare two outcomes: first, the fair chances option is chosen and the coin comes up tails, so Talia gets the kitten; second, you decide to give the kitten to Talia, and so you do. At the end of the day, has everything turned out the same in the two schemes? Are the outcomes the same? In one sense, obviously, they are: the outcome of each is that Talia gets the kitten. But in the second scheme Horace has a complaint, and in the first he does not. Isn't that a difference? In the second scheme, Horace has been treated unfairly. In the first scheme, there is no unfairness. That seems to be a difference, and it is no strain to say that it is a difference in the outcomes. Indeed, we might say, if everything really turns out just the same in the end, then neither case could be better than the other. The point is that we may have been too quick in identifying outcomes. The outcome of the fair chances option is that Talia gets a kitten fairly, while the outcome of giving the kitten to Talia outright is that she gets a kitten unfairly. A proper representation of the two options ought to represent that difference. So maybe they should be represented like this: L[1 2 , t u , t u ] and L[1 2 , t f , h] But then the example does not even appear to violate better prizes. The issue is crucial to Harsanyi's argument. It shows up again in a slightly different guise in the question of the soundness of an extra premise we need for that argument. To recap, we are going to use the assumption that at least as good as for persons satisfies the decision theory axioms, and the assumption that M satisfies the axioms, and then two more premises. The first of these we can call Personal Good: 30 end p.174 (PG)If x is at least as good for each person as y, and better for some person, then x is morally better than y; and if x is exactly as good for each person as y, then x is exactly as good as y.
We need one more premise. Following standard practice I will call it Anonymity: (A) Nobody's good matters more than anyone else's. Surveying the distribution of good to persons, according to (A), we ought to be able to see how (morally) good the situation is without knowing the identities of the bearers of good. The persons can be just as well left anonymous.
4. Harsanyi's Theorem John Harsanyi put his theorem in terms of the more standard interpretation of decision theory. He imagined that the preferences of individual members of a society met the constraints of decision theory, and that a “social preference” ordering did too. Adding the analog of (PG) and (A) to connect the members' preferences with the social ordering, he showed that the social preference ordering must be representable by a utility function that is the sum of the utility functions of the individual members. But the conclusion of the argument is more obviously relevant to ethics if it is about what is morally good and what is good for individual people. So as we will interpret the theorem, it says that there is a utility function that represents moral goodness and is the sum of utility functions representing the good of persons. 31 Step 1: Choice of Scales Since we have leeway in our choice of utility functions (we may choose from the family related by positive linear transformation), we will choose some that are convenient for the presentation. So first, we will represent the worst prospect for any given person by the utility number 0, and the best by the utility number 1. 32 Let U 1 , U 2 , etc. be the utility functions representing the good for persons 1, 2, etc. And the utility function representing overall moral goodness will be W. W must assign the same number to any pair of alternatives that are assigned the same number by all of the U i . 33 Also, (PG) tells us that the alternative in which every end p.175 person is worst off must be the worst possible alternative, since every distinct alternative will be better for someone and worse for no one. We'll select our moral goodness scale so that this morally worst alternative is assigned 0. This means that W(0, 0, 0,, 0) = 0 where the arguments are, as I said, the utility numbers measuring the good for each person. We complete the choice of a scale for W by stipulating that W(1, 0, 0,, 0) = 1
By (A), W(1, 0, 0,, 0) = W(0, 1, 0,, 0) == W(0, 0, 0,, 1) = 1 That is, all the alternatives that are good to degree 1 for one person and worst for each other are exactly as good as one another (that's Anonymity), and we've stipulated that the degree of moral goodness of these alternatives is one. We'll call these “unit vectors.” 34 Step 2: Multiplication Lemma What I'll call the multiplication lemma says that for any real number k between 0 and 1, (ML) kW(u 1 , u 2 ,, u n ) = W(ku 1 , ku 2 ,, ku n ) where the u i are the utility numbers representing how good the alternatives are for persons 1 through n. That is to say, halving the goodness for each person will half the overall goodness, and generally multiplying the goodness for each person by k will multiply the overall goodness by k. This is hardly obvious. But it is not hard to demonstrate. Consider the lottery L* = L[k, (u 1 , u 2 ,, u n ), (0, 0,, 0)] Because all of the utility functions are expectational, we can work out quickly some definite facts about the utilities of L*. For example, the goodness of L* for person 1 must be ku 1 , since that is the expected goodness for person 1. U 1 (L*) = ku 1 ; U 2 (L*) = ku 2 ;; U n (L*) = ku n The overall goodness of the worst case scenario is also 0, so the expected moral goodness of L* is also a simple matter: W(L*) = kW(u 1 , u 2 ,, u n ) But now we may just replace the “L*” on the left side with the list of utilities representing the goodness of L* for each person: W(ku 1 , ku 2 ,, ku n ) = kW(u 1 , u 2 ,, u n ) which is the multiplication lemma. Step 3: Harsanyi's Theorem We will show that W(u 1 , u 2 ,, u n ) = u 1 + u 2 ++ u n , that is, that the moral goodness is the sum of the goodness for persons. The proof will be for the case of two persons only; the generalization to many persons is direct but because of our notation it would be very cumbersome. If we are given u 1 and u 2 , we can then consider the lottery L[1/2, (u 1 , 0), (0, u 2 )] giving equal chances to (u 1 , 0) and (0, u 2 ). Call this lottery L= . Its moral goodness must be equal to its expected moral goodness, so W(L= ) = W(u 1 , 0)/2 + W(0, u 2 )/2 Its expected goodness for person 1 is u 1/ 2, and for person 2 u 2/ 2, so U 1 (L= ) = u 1/ 2; U 2 (L= ) = u 2/ 2 The moral goodness of L= is just a matter of how good it is for each person, so W(L= ) = W(u 1/ 2, u 2/ 2), so by the multiplication lemma, W(L= ) = W(u 1 , u 2 )/2
Now combine the first and last lines to get W(u 1 , 0)/2 + W(0, u 2 )/2 = W(u 1 , u 2 )/2 end p.177 so as a matter of algebra, W(u 1 , 0) + W(0, u 2 ) = W(u 1 , u 2 ) But the multiplication lemma tells us that W(u 1 , 0) + W(0, u 2 ) = u 1 + u 2 So, W(u 1 , u 2 ) = u 1 + u 2 , which is what we wanted to show. Remarks on Harsanyi's Theorem The conclusion of the argument rests on a handful of assumptions. Several of these assumptions are questionable. However, the argument is still remarkable, I think. The assumptions appear to be a lot weaker than the conclusion (though of course, collectively they couldn't be weaker than the conclusion!). I want to point out two things. First, we have to revisit Anonymity. Second, I want to elaborate on a crucial step in the argument, the third step, explaining how one of the controversial assumptions figures in that step. First, for Anonymity to do the work it has to do in the proof, we must in effect assume that we have a clear way to compare how well off I am with how well off you are. If we have such a thing, then Anonymity really does mean that swapping numbers from one position to another on a vector cannot make it better or worse. But if the goodness scales for you and me are really just arbitrary, so that your 0 and my 0 are in no sense “the same degree of welfare,” then it is hard to see what justification we could have for supposing that permuted vectors must be exactly as (morally) good as one another. The assumption that we can compare my welfare with yours is a strong one. It does seem reasonably intuitive that we can do it, but it is hard to see how to do it rigorously. Now to (what I think is) the crucial step, the third one. There we considered a fair lottery between (u 1 , 0) and (0, u 2 ) and said that this lottery must have goodness of u 1/ 2 for person 1 and u 2/ 2 for person 2. We might plug in numbers to make things clearer; let's have u 1 = u 2 = 2, to fix ideas. Then these lotteries are good for person 1 (me, say) to degree 1 and for you (person 2) also to degree 1. That means the lottery is exactly as good for you as (1, 1) and also exactly as good for me as (1, 1). Call the (1, 1) outcome the egalitarian outcome, and (2, 0) and (0, 2) the inegalitarian outcomes. Since the lottery between the inegalitarian outcomes is exactly as good for each of us as the egalitarian outcome is, (PG) implies that the lottery is exactly as good morally as the egali end p.178 tarian outcome. But one might think the egalitarian outcome is better. After all, it is egalitarian! And that is something that can't be said for the outcomes of the lottery. Still, it might be thought that an egalitarian outcome is not better than a fair lottery between inegalitarian outcomes. That person 1 does better than person 2 may not be an objection, as long as person 2 had an equal chance of doing better. The fairness may reside in the whole of the lottery, as it did in the kitten story. But notice now that the outcome (0, 2) must be exactly as good as the outcome (2, 0). But then better prizes tells us that (0, 2) is exactly as good as L[1/2, (2, 0), (0, 2)]. This seems
wrong, and for exactly the reason that giving the kitten to Talia seemed intuitively worse than holding a fair lottery between Talia and Horace. And again, better prizes is the culprit. Now as before, we might say that the outcomes are not properly represented. If person 2 ends up on the short side of things as the result of a fair lottery, that is not the same outcome as her having had no chance of the better outcome. But here the outcomes are represented only by a sequence of utility numbers. If these numbers do represent goodness for the persons involved, and if fairness counts extra toward the overall goodness of the outcome, then that extra goodness couldn't be goodness for some person. That means that (PG) is false. Maybe fairness is a good that is not a good for anybody. The argument of Harsanyi's theorem makes some strong assumptions about where goodness of an outcome can be located. It assumes that all goodness must be goodness for someone (that's what (PG) says), and also that the goodness of a prospect must reside entirely in the goodness of the prospective outcomes (that's what better prizes says). These are dominance principles, and also “antiholism” principles. Goodness of a complex must reside in the parts, and not in their relations to one another. These assumptions may be wrong. Even if they are wrong, I think Harsanyi's theorem provides an illuminating argument. It gives us one way of thinking about which assumptions are fundamental to utilitarianism. And it shows how these assumptions figure in the overall theory. Considerations of better prizes and (PG) may also show something about utilitarianism and may help to explain certain counterintuitive features of the view. Utilitarianism does not recognize the good of equality; it does not recognize how distributive considerations may be important. Its antiholism is traceable to the dominance principles at work in the Harsanyi argument. end p.179
NOTES 1. I'll follow the terminology of Resnik 1987 , which follows von Neumann and Morgenstern's formulation. 2. See Dreier 1996 for a defense of this interpretation. 3. See Resnik 1987 , chap 4. Resnik's book provides compact and clear explanations of both formal and philosophical topics within and arising out of decision theory. 4. It is proved in Resnik 1987 , 88–98. 5. In this paragraph, and hereafter, I use the capitalized “Cooperate” and “Defect” as names for moves in the game of Prisoners' Dilemma. 6. The structure of the two games depends only on the ordering of the outcomes for each player. I have selected the payoffs in such a way that the total utility is greater for the two players combined in the perfectly cooperative outcome, and least in the perfectly uncooperative outcome. This feature will help me make a couple of points later. 7. Beginning with Rawls 1951 , and fading off rather than ending at a sharp boundary line, but in any case most famously in Rawls 1971 . 8. In a number of papers, but culminating in Gauthier 1986 .
9. The theory is Kantian in another respect. Morality is binding on us not merely in virtue of the preferences we happen to have, not merely because or insofar as we happen to care about others, but also in virtue of our rationality. 10. In Hobbes 1968 , chap. 15. 11. Gauthier 1986 , 171. There is another clause, which I omit for brevity. 12. Axelrod 1984 gives a very satisfying and compelling presentation of some computer simulations that demonstrate the superiority, in many different environments, of reciprocal cooperation. 13. Some people think that Newcomb's Predictor acts in exactly this way. See Lewis 1985 . 14. See Gauthier 1986 , chap. 5, especially 136–45. 15. In fact, it bucks the tradition of Nash, Zeuthen, and Harsanyi, who argue for a different method of determining the rational bargain. See Luce and Raiffa 1989 , 124ff. 16. This argument is not easy to find. Here I am working from Gauthier 1986 , 137–39. 17. In this paragraph I have not capitalized “cooperate,” “defect,” or their cognates. That's because the capitalized terms are technical terms; they are the names of moves in the theoretic game of PD, as I explained in n. 5. 18. See Blackburn 1998 , chap. 6, for a similar diagnosis of what has happened when somebody becomes a constrained maximizer. 19. See Gauthier's many discussions of “mutual unconcern” and “non-tuism.” Gauthier does not construe his own argument as I am suggesting here. For an interesting discussion of Gauthier's own use of the assumption of “mutual unconcern,” see Thomas 1988 . 20. Her attitude toward her own cooperative attitudes might be something like the attitude that Hume ([1739 ] 1978 ) thought we should have toward our “artificial virtues,” in book 3, part 2 of the Treatise. 21. See Locke's Second Treatise of Government ([1690a ] 1952 ), chap. 5, paras. 27 and end p.180 33: “Whatsoever [a man] removes out of the state that nature has provided, he has mixed his labor with, and joined to it something that is his own, and thereby makes it his property.For this labor being the unquestionable property of the laborer, no man but he can have a right to what that is once joined to, at least where there is enough and as good left in common for others.” The Lockean Proviso, so-called by Robert Nozick (1974 , 175ff.), is the portion italicized. 22. Or at least it approximates that idea. See Hubin and Lambeth 1991 for a discussion of complications. 23. This claim may seem dubious. It is discussed by several critics in part 2 of Vallentyne 1991 . 24. Vallentyne 1991 is a collection of uniformly high quality and remarkably broad interest. 25. See, e.g., Temkin 1996 . 26. As explained in Hausner 1954 . 27. In Allais 1953 . 28. See, e.g., Thomson 1997 and Foot 1988 .
29. The example is modeled after one given in Diamond 1967 . 30. Following Broome 1991 , especially 165. Broome's book provides a superb philosophical treatment of the application of decision theory to the ethical concept of goodness, in far more depth than I manage in this chapter. 31. Aside from this difference in interpretation and a few minor differences, the proof here follows Resnik 1987 , 197–200. 32. This assumes that there are best and worst possibilities for persons, which may not be true. The proof still works without the assumption, but it is more cumbersome. 33. For this reason, I am going to use the functor W indiscriminately as a functor of one place, namely the option to be morally evaluated, and a functor of many places, namely the utilities representing how good the option is for each person. I hope this reduces confusion rather than aggravating it. 34. It is sometimes asked whether we can be sure all the unit vectors exist. The question is a little more worrisome in the context of Harsanyi's original argument, where the locations on the vectors represent people's preferences. In our context, there could be a real worry if we insisted that each alternative be actual, but there is no particular reason so to insist.
Chapter 10 RATIONALITY AND GAME THEORY Cristina. Bicchieri Game theory (see also Danielson, chap. 22, this volume) aims to understand situations in which decision makers interact. Chess is an example, as are firms competing for business, politicians competing for votes, jury members deciding on a verdict, animals fighting over prey, bidders competing in auctions, threats and punishments in long-term relationships, and so on. What all these situations have in common is that the outcome of the interaction depends on what the parties jointly do. Decision makers may be people, organizations, animals, robots, or even genes. The theory of rational choice is a basic component of game-theoretic models. This theory has been mostly criticized from a descriptive viewpoint, arguing that it requires far too many calculating capabilities from ordinary beings that use at most simple heuristics. Few, however, have given a close and critical look at how a normative theory of rational choice fares in interactive decision contexts. What does it mean to be rational when the outcome of one's action depends upon the actions of other people and everyone is trying to guess what the others will do? In social interaction, rationality has to be enriched with further assumptions about individuals' mutual knowledge and beliefs, but these assumptions are not without consequence. In what follows I shall spell out these extra assumptions, and see whether they are sufficient to lead the players to coordinate upon outcomes that are mutually acceptable. Since the issue of whether rational choice theory is an adequate foundation for game theory is raised in the end p.182
context of noncooperative games, I will restrict my attention to this class of games. Rational Choice The theory of rational choice's central assumption is that a decision maker chooses the best action available according to her preferences. The content of preferences is unrestricted. Agents' preferences may be selfish or altruistic, self-defeating or even masoch*stic. Preferences mirror values and dispositions that are beyond the pale of rationality. What is required is that preferences are well behaved, in the sense of fulfilling certain formal conditions I list below. If preferences are well behaved, they can be represented by utility functions, and rationality consists in maximizing one's utility, or finding the maximum value of one's utility function. When we model a choice situation, we must be careful to follow a series of steps. First, we must define a set of feasible actions A. These are the actions an agent knows are available to him. If an action is available, but one does not know about it, then we do not include it in the set of feasible actions. In the simplest case, each action has a unique, certain consequence. We must then list the set C of the consequences of actions. Since an agent is assumed to have preferences over consequences, the third step involves introducing a binary weak preference relation R on the set C that is complete (for any x, yC either xRy or yRx), antisymmetric (if xRy and yRx, then x y), reflexive (for all xC, xRx) and transitive (if x, y, z C, xRy and yRz, then xRz). ( denotes indifference, P (used below) denotes strict preference.) The preceding assumptions about the weak preference relation on C identify a qualitative preference structure. A representation provides a correspondence between this structure and properties of real valued functions based on C. In other words, we want a numerical representation for the preference relation, such that higher numbers correspond to more preferred consequences. Utility functions are just such representations. An agent's preferences will then be represented by a utility function U: C → R , such that xRy iff U(x) ;eg U(y), for any x, yC. Since in this case the utility function only conveys ordinal information, any increasing function of U can represent an agent's preferences. The final step consists in defining rational choice proper. If each action has a unique consequence, we may introduce a consequence function f: A→ C that associates a consequence with each action. A rational agent will choose an action a* that is feasible and optimal, in the sense that, for all aA, f(a*) R end p.183 f(a). Alternatively, we may say that a rational agent will act as if maximizing the utility function U(f(a)). A far more common case is one in which an action may have one of several possible consequences. A situation of risk is one in which the probabilities with which the consequences occur are objective and known to the decision maker. Then the consequence function f is stochastic and is known to the agent (i.e., for each action a, f(a) is a probability distribution on C). Suppose action a has two possible consequences, x and x, which occur with probability p and (1–p), respectively. Choosing action a is like choosing a lottery that gives prize x with probability p, and prize x with probability (1–p). We assume agents to have preferences over such lotteries. If preferences are complete, transitive, and satisfy a number of other conditions (von Neumann and Morgenstern 1944 ), they can be represented by the expectation of a real-valued utility function U: C → R
(unique up to a positive linear transformation) such that, for any two lotteries a and b, aPb iff Σ xa p(x)U(x) > Σ yb p(y)U(y). A rational agent will choose an action (lottery) a* that maximizes the expected value of a von Neumann–Morgenstern utility function. If the stochastic connection between actions and consequences is not given, the decision maker will have to make some hypotheses about which states of nature might obtain, since the consequence of an action will depend upon which of a number of mutually incompatible states of nature is realized. The ingredients of choice under uncertainty are thus consequences and states of nature. A state of nature indicates, for each feasible act, what the consequence will be. The consequence function f is now defined as f: A X S →C, so when we write f(a, s), we mean the consequence of taking action a if state s is realized. The desirability of the consequence depends neither on the act, nor on the state of nature that makes it possible. An agent is assumed to consider a state space S, and to act in a manner consistent with having a subjective probability measure p over S (Savage 1954 ). p is an additive subjective probability measure over S. A subjective expected utility representation for a system of preferences exists if preferences satisfy some necessary conditions formulated by Savage (1954 ). In this case, there exists a real-valued utility function U: C → R such that, for any two actions a and b, aPb iff Σ sS p (s)U(f(a, s)) > Σ sS p (s)U(f(b, s)). A rational agent will select an action a* that maximizes the expected value of U(f (a*, s)), relative to the probability measure. In all three cases, rationality is identified with (expected) utility maximization, and this presupposes that individuals have (or behave as if they have) utility functions, that is, that they have well-behaved preference orderings over alternatives. In game-theoretic models, von Neumann–Morgenstern utility functions are the most commonly used. Such utility functions are unique only up to a positive linear transformation, which means that for any set of outcomes there is an in end p.184 finite number of numerical functions, mutually related by positive linear transformations, each of which represents an agent's utility as well as any other. A consequence of using von Neumann–Morgenstern utility functions is that it becomes impossible to make interpersonal utility comparisons. To understand why, imagine trying to compare two temperatures that are expressed in Celsius and Fahrenheit measures, without having a scale. In noncooperative games this is not a problem, since interpersonal utility comparisons are never brought into play. In evolutionary games instead, interpersonal utility comparisons may be needed. This is especially true in models of cultural evolution, the process through which behavioral patterns that are not genetically inherited are transmitted from one generation to another, or spread from one group to other groups. One such transmission process is imitation, which might involve comparing one's payoffs with another's, and switching strategies if the other's payoffs are better than one's own. Interpersonal comparisons may be a completely subjective matter, but if this were the case, we could not talk of a selection process that favors better, more efficient strategies over less efficient ones. For selection to work, we need at least intersubjective agreement on which strategies are “most successful,” and this involves interpersonal comparisons among payoffs. If payoffs are expressed in von Neumann–Morgenstern's utilities, however, such comparisons are impossible.
Strategic Interaction In a strategic interaction, the outcome of an action depends, among other things, upon the actions of other agents. Other agents have plans, preferences, and beliefs, and unless one is certain which action will be chosen by another agent, one will have to form beliefs about other agents' possible choices, and even beliefs about the expectations that may guide another agent in choosing a particular action. Whereas rational choice is relatively straightforward in individual decision making, it becomes more complicated in a strategic decision context. When we abstract from the particulars of such contexts and formally describe the strategic interaction, we have a game. Any strategic interaction involves two or more decision makers (players), each with two or more ways of acting (strategies), such that the outcome depends on the strategy choices of all the players. Each player has well defined preferences among all the possible outcomes, enabling corresponding von Neumann–Morgenstern utilities (payoffs) to be assigned. A game makes explicit the rules gov end p.185 erning players' interaction, the players' feasible strategies, and their preferences over outcomes. One way of representing games is in normal form. A normal form game is completely defined by three elements: a list of players i = 1,, n; for each player i, a finite set of pure strategies S i ; a payoff function u i that gives player i's payoff u i (s) for each n-tuple of strategies (s 1 ,, s n ), where u i : X>[md4]n[mu4][mu6]>i=1[md6] Si → R . All players other than some given player i are customarily denoted as −i. A player may choose to play a pure strategy, or instead he may choose to randomize over his pure strategies; a probability distribution over pure strategies is called a mixed strategy and is denoted by σ i . The pure strategies over which a player randomizes are called the support of the resulting mixed strategy. Each player's randomization is assumed to be statistically independent of that of his opponents, and the payoffs to a mixed strategy are the expected values of the corresponding pure strategy payoffs. The 2×2 matrix in figure 10.1 depicts a two-player normal form game: each player picks a strategy independently, and the outcome, represented in terms of players' payoffs, is the joint product of these two strategies. Often the normal form representation is interpreted as a situation in which the players choose simultaneously, without any knowledge of each other's choice. This temporal interpretation, though intuitive, is not quite correct. Since the normal form is not meant to convey any information about players' knowledge of each other's moves, or their temporal order, it is best to interpret it as a list of outcomes, one for each possible combination of strategies the players might choose. If we want to model the information (or lack thereof) that players have about each other and the order in which they move, we have to choose another type of representation. I will return to this point later.
Figure 10.1 The game of figure 10.1 is one of complete information, in that players are assumed to know the rules of the game (which include players' strategies) and other players' payoffs. If players are allowed to enter into binding agreements before the game is played, we say that the game is cooperative. Noncooperative games instead make no allowance for the existence of an enforcement mechanism that would make the terms of the agreement binding on the players. If preplay negotiation leads to an agreement, and there is no enforcement mechanism available, we should ask whether the agreement is selfenforcing, by virtue of being in the best interest of each player to adhere to it.
Nash Equilibrium Expected utility maximization has always been a building block of game theory, but for many decades game theorists have paid little attention to the link between rational choice and strategic interaction, or how the outcome of strategic interaction can be derived from rational choices. In a well-known passage of their book, Theory of Games and Economic Behavior, von Neumann and Morgenstern say that rational players who know (i) all there is to know about the structure of the game they are playing, (ii) all there is to know about the beliefs and motives of the other players, (iii) that every player is rational, (iv) that every player knows (i)–(iii), (v) that every player knows (i)–(iv), and so on, will be able to infer the optimal strategy for every player. In that case, each player will behave rationally by maximizing his expected utility conditional on what he expects the others to do. This idea that rational players will always jointly maximize their expected utilities, or play a Nash equilibrium, could rightly be called the “central dogma” of game theory. Nash equilibrium (Nash 1951 ) is the standard solution concept for noncooperative games. Informally, a Nash equilibrium specifies players' actions and beliefs such that (i)
each player's action is optimal given his beliefs about other players' choices, and (ii) players' beliefs are correct. Thus an outcome that is not a Nash equilibrium requires either that a player chooses a suboptimal strategy, or that some players “misperceive” the situation. More formally, a Nash equilibrium is a vector of strategies (σ* 1 ,, σ* n ), one for each of the n players in the game, such that each σ* i is optimal given (or is a best reply to) σ* −i . Note that optimality is conditional only on a fixed σ −i , not on all possible σ −i . A strategy that is a best reply to a given combination of the opponents' strategies may fare poorly vis-à-vis another strategy combination. Consider for example figure 10.2 . end p.187
Figure 10.2 In this game, there are two Nash equilibria, (C, C) and (D, D), but a player's strategy D, for example, is optimal only if the other player is expected to play D, too. If the other player were to play C instead, the choice of D would lead to a poor outcome. A common interpretation of Nash equilibrium is that of a self-enforcing agreement. Were players to agree in preplay negotiation to play a particular strategy combination, they would have an incentive to stick to the agreement only in case the agreed upon combination is a Nash equilibrium. In the case of a strict Nash equilibrium, any deviation from the equilibrium strategy nets a player an inferior payoff. If the equilibrium is not strict, however, a deviation from equilibrium play may earn a player the same payoff as the equilibrium strategy. In the latter case, the incentive to follow the Nash equilibrium is less strong. The lack of a strong incentive to play one's part in a Nash equilibrium is particularly obvious in the case of mixed strategy equilibria, which, as we shall see, are never strict. Consider the game in figure 10.3 . This game has no Nash equilibrium in pure strategies, but Nash proved that—provided certain restrictions are imposed on strategy sets and payoff functions—a game has at least an equilibrium in mixed strategies. Nash's result generalizes von Neumann's theorem
(1928 ) that every noncooperative game with finitely many strategies has an equilibrium in mixed strategies. Suppose 1 plays (4/9 a, 5/9 b). Then if 2 chooses c, her expected utility is 4(4/9) + 7(5/9)= 17/3. If 2 chooses d, she nets 9(4/9) + 3(5/9) = 17/3. So if 1 randomizes between a and b with probabilities (4/9, 5/9), 2 is indifferent between c, d, or a lottery in which she chooses c with probability p and d with probability (1–p). Suppose 2 chooses (4/7 c, 3/7 d). In this case 1 nets 48/7 if he plays a, and 48/7 if he plays b. Hence 1 is indifferent between a, b, and any lottery (ap, b(1–p)). The combination (4/9 a, 5/9 b), (4/7 c, 3/7 d) is a mixed strategy Nash equilibrium. end p.188
Figure 10.3 In a mixed strategy equilibrium, the equilibrium strategy of each player makes the other indifferent between the strategies on which he is randomizing. For example, if 1 were to know that 2 randomizes with probabilities (4/7, 3/7), any of his strategies (pure or mixed) would be a best reply to 2's choice, and conversely, were 2 to know that 1 randomizes with probabilities (4/9, 5/9), any of her strategies, pure or mixed, would be a best reply. Paradoxically, if players agree to play a mixed strategy equilibrium, they have no incentive to play their part in the equilibrium. A mixed strategy equilibrium is a selfenforcing agreement only in the weak sense that—given the other players' equilibrium behavior—each player is indifferent between all the strategies (and lotteries over these strategies) in the support of her equilibrium mixed strategy. There are, however, more serious questions raised by the Nash equilibrium concept. Ken Binmore (1987 , 1988 ) has argued that there are two possible interpretations of Nash equilibrium. According to the evolutive interpretation, a Nash equilibrium is an observed regularity. Players know the equilibrium, and test the rationality of their behavior given this knowledge acquired from experience. The players (and the game theorist) can accordingly predict that a given equilibrium will be played, since they are accustomed to
coordinate upon that equilibrium and expect (correctly) others to do the same. According to the more commonly adopted eductive interpretation instead, a game is a unique event. In this case it makes sense to ask whether players can deduce what others will do from the information available to them. The players (and the game theorist) can predict that an equilibrium will be played just in case they have enough information to infer players' choices. The standard assumptions game theorists make about players' rationality and knowledge should in principle be sufficient to guarantee that an equilibrium will obtain. The following assumptions are standard: end p.189 CK1. The structure of the game, including players' strategy sets and payoff functions, is common knowledge among players. CK2. The players are rational (i.e., they are expected utility maximizers) and this is common knowledge. The concept of common knowledge was introduced by Lewis (1969 ), and later formalized by Aumann (1976 ). Simply stated, common knowledge of p among a group G means that each member of G knows p, and each knows that each knows p, and so on ad infinitum. Common knowledge of rationality, preferences and strategies may facilitate the task of predicting an opponent's strategy but, as I argued elsewhere (Bicchieri 1993 ), it does not guarantee that the resulting prediction will be correct. Consider a game that has a unique Nash equilibrium in pure strategies (figure 10.4 ). Can the players infer what other players will do from CK1 and CK2? Here player 1 has two pure strategies, A and B, and player 2 has three pure strategies, a, b, and c. There is a unique Nash equilibrium in pure strategies, (B, a), but it is not evident that players can infer that it will be played by reasoning from CK1 and CK2. As an example of how players may reach a conclusion on how to play, consider the following argument by player 1. “If player 2 believes that I will play A, then it is optimal for her to pick c. And why would she think I play A? Well, she must believe that I expect her to play b, to which A is a best reply. And why would I expect her to play b? I would (she will think), if I were to believe she expects me to play B” It is easy to verify that such a chain of reasoning can justify the choice of any strategy for both players. The concept of Nash equilibrium embodies a notion of individual rationality, since each player's equilibrium strategy is a best reply to the opponents' strategies, but unfortunately it does not specify how players come to form the beliefs about each other's strategies that support equilibrium play. Beliefs, that is, can be internally consistent but fail to achieve the interpersonal consistency that guarantees that an equilibrium will be attained. Bernheim (1984 ) and Pearce (1984 ) have argued that assuming players' rationality (and common knowledge thereof) can guarantee only that a strategy will be rationalizable, in the sense of being supported by internally consistent beliefs about other players' choices and beliefs. But a combination of rationalizable strategies may not constitute a Nash equilibrium. In the game depicted in figure 10.4, all six combinations of strategies are rationalizable, yet only one of them is an equilibrium. The fact that a Nash equilibrium is
always a combination of rationalizable strategies is of no help in predicting it will be played. Note that there are strategies that are not rationalizable, in the sense of not being supported by coherent beliefs. Consider the game in figure 10.1. Here the C-strategy is never rationalizable, since each player is always better off by choosing end p.190
Figure 10.4 strategy D, whatever the other does. And each player is able to see that the other player, if rational, will never choose strategy C. In this case, CK1 and CK2 will lead the players to predict the outcome of the game. In a Nash equilibrium, the optimality of a strategy is conditional only on a fixed σ −i , not on all possible combinations σ −i . In the game of figure 10.1, however, the Nash equilibrium strategies (D, D) are also optimal with respect to any strategy choice of the opponent. Whatever player 1 does, player 2 is better off by choosing D, and the same is true of player 1. We say that a strategy s i is strictly dominated by another strategy t i if, for every choice of strategies of the other players, i's payoff from choosing t i is strictly greater than his payoff from choosing s i . In our example, C is strictly dominated by D for both players. A strictly dominated strategy is never rationalizable, since the belief that a player plays it is inconsistent with common knowledge of rationality. We say that s i is weakly dominated by t i if, for every choice of strategies of the other players, i's payoff from choosing t i is at least as great as i's payoff from choosing s i . Note that weakly dominated strategies are rationalizable, since there always exists an opponents' strategy combination to which a player's weakly dominated strategy is a best reply. When a game has a unique Nash equilibrium, we can predict that it will be played if we are able to show that players, armed with common knowledge of rationality and of the structure of the game, will infer the Nash solution. If players have dominated strategies, CK2 entails that they will eliminate them, and this is common knowledge (we assume
that the consequences of CK1 and CK2 are common knowledge, too). Often after we have eliminated strictly dominated strategies for one player, we may find that there are now strictly dominated strategies for another player, which will be eliminated as well. This process of successive elimination can continue until there are no more strictly dominated strategies left. If a unique strategy remains for each player, we say the game has been solved by iterated dominance. It is easy to prove that a strategy profile thus obtained is a Nash equilibrium (Bicchieri 1993 ). Consider for example the game in figure 10.5 . R is a strictly dominated strategy for player 2, and since rationality is common knowledge, 2 is expected to eliminate R as a possible choice. Player 1 will now expect L to be played, in which case U dominates D. (U, L) is the solution to the game, and it is inferrable from CK1 and CK2. Note that assuming common knowledge of rationality (or at least some level of mutual knowledge of rationality) is crucial to obtaining the (U, L) solution. If there were some doubt about a player's rationality, the solution would unravel. For example, if 1 were to think there is a 0.01 chance that R is chosen, then he would be better off by choosing D. In real life this is likely to occur. That is, in real life a player may “play safe” and prudently choose D, but we are now discussing a completely different point. The question is not whether players are fully rational or believe each other to be. Rather, we want to know how far they can go in inferring a Nash solution from CK1 and CK2. As we have seen, the answer often is not far. Predictability is hampered by another common problem encountered in game theory: multiple Nash equilibria. Suppose two players have to divide $100 among themselves. They must restrict their proposals to integers, and each must propose a way to split without knowing the other's proposal. If the total proposed by both is equal or less than $100, each gets what she proposed, otherwise they get nothing. This game has 101 Nash equilibria. Is there a way to predict which one will be chosen? Alternatively, is there a way a player can infer what the other will do, and thus adjust her proposal accordingly? In real life, many people would go for the 50/50 split. It is simple, and seems equitable. In Schelling's words, it is a focal point (Schelling 1960 ). A focal point equilibrium has some property that makes it salient. A solution may be salient because of historical precedent, or because it embodies cultural norms we share (Lewis 1969 ). Unfortunately, mere salience is not enough to provide a player with a reason for choice. In our example, only if it is common knowledge that the 50/50 split is the salient outcome does it become rational to propose $50. Game theory, however, filters out any social or cultural information regarding strategies, leaving players with the task of coordinating their actions on the sole basis of common knowledge of rationality (and of the structure of the game). Consider now another game that many readers would intuitively know how to solve. The game of figure 10.6 has two Nash equilibria in pure strategies: (C, C) and (D, D), but in the (D, D) equilibrium each player plays a weakly dominated strategy. (C, C) is a Pareto-dominant equilibrium point, since it gives both players a higher payoff than any other equilibrium in the game. For this very reason, it should be a natural focal point for both players. This game is an example of a end p.192
Figure 10.5 coordination game, that is, a game that has several Nash equilibria such that it is always preferable for a player to coordinate with the other and play one of the equilibria, since lack of coordination yields each player a lower payoff. Note that in the above example players' interests coincide, but even if we change the outcomes of (C, C) and (D, D) in figure 10.6 to (3, 1) and (1, 3), respectively, it remains true that both players prefer to coordinate upon one of the two equilibria to “going solo,” even if each prefers a different equilibrium. Should we confidently predict that (C, C) will be the solution of the game in figure 10.6? We have seen that focal points must be common knowledge among the players before it becomes rational for them to play the focal point equilibrium. If no such common knowledge is present, rationality alone is not a reliable guide. Could elimination of weakly dominated strategies do the trick? We know that
Figure 10.6 end p.193 when a player has a strictly dominated strategy, rationality dictates eliminating it, hence predicting behavior is a (relatively) simple matter. The case of weakly dominated strategies is not that straightforward. For one, a weakly dominated strategy is still a best reply to some opponent's strategy. Putting it differently, weak dominance means that there is at least one choice on the part of an opponent that makes one indifferent between the weakly dominated strategy and some other strategy. In our example, were player 2 to believe that 1 plays D, she would be indifferent between C and D, since both C and D are best replies to D, and conversely, were player 1 to expect 2 to play D, he would be indifferent between C and D, since both strategies are best replies to D. One possible solution is to introduce a rule according to which weakly dominated strategies should also be eliminated by a rational player. The problem is that a player may be indifferent between two strategies, one of which is weakly dominated by the other, if she treats as null the state on which the weakly dominant strategy is strictly preferred. To guarantee that a player will always eliminate a weakly dominated strategy, we must assume that no state of the world is treated as null by the players. This means that a player's full belief that a strategy will be played is not to be interpreted as treating the event that it won't be played as null. As some philosophers have argued (Harper 1976 , McGee 1994 ), a state that is not considered an epistemically serious possibility need not be treated as null. This means that a player's conditional preference for a strategy, given such an “impossible” state, can be nontrivially defined. A related approach using lexicographic probabilities to reconcile iterated elimination of weakly dominated strategies and Bayesian decision theory is taken in Blume, Brandenburger, and Dekel 1991 and Stahl 1995 . Another difficulty with applying iterated elimination of weakly dominated strategies is that the order of elimination usually matters; thus a particular solution will depend on which strategy was first eliminated. A standard solution to this problem is to delete at each round all weakly dominated strategies of all players (Rochet 1980 , Moulin 1986 , Harper 1991 , Bicchieri and Schulte 1997 ). Eliminating weakly dominated strategies is an example of an “eductive” procedure. When asking how the players' deductive processes might unfold, one must specify some basic principles of rationality, and then examine which choices are consistent with common knowledge of the specified principles. Such choices may or may not result in an equilibrium, but at least the link between rational choice and equilibrium (when there is such link) is made clear. The advantage of this approach is that it is possible to refine our predictions about how players might choose without assuming that they will coordinate on a particular equilibrium. Principles such as iterated strict dominance and rationalizability are examples of how it is possible to restrict the set of predictions using rationality arguments alone. In most cases, however, the set of possible outcomes is still too large. end p.194
A very different approach to the problem of indeterminacy is to start by considering the set of Nash equilibria, and to ask whether some of them should be eliminated because they are in some sense “unreasonable.” This is the approach taken by the refinement program (Kohlberg 1990 ; van Damme 1991).
Normal Form Refinements Consider again the game in figure 10.6. How reasonable is the equilibrium (D, D)? Under what circ*mstances would players agree to play it, and then stand by the agreement? The equilibrium strategies (D, D) are weakly dominated but—as I already argued—common knowledge of rationality does not force players to eliminate them. Prudence, however, may suggest that one should never be too sure of the opponents' choices. Even if players have agreed to play a given equilibrium, some uncertainty remains. If so, we should try to model this uncertainty in the game. Selten's insight was to treat perfect rationality as a limit case (Selten 1965 ). His “trembling hand” metaphor presupposes that deciding and acting are two separate processes, in that even if one decides to take a particular action, one may end up doing something else by mistake. An equilibrium strategy should be optimal not only against the opponents' strategies, but also against some very small probability > 0 that the opponents make “mistakes.” Such an equilibrium is tremblinghand perfect. Is the equilibrium (D, D) perfect? If so, player 2's choice of D must be optimal against C being played (by player 1) with probability and D being played (by player 1) with probability 1− for some small > 0. But in this case player 2's payoff to C is 3 , whereas his payoff to D is . Hence for all > 0, C is a better strategy choice for player 2. Since the game is symmetrical, the same reasoning applies to player 1. The equilibrium (D, D) is not (trembling-hand) perfect, but (C, C) is. A prudent player therefore would discard (D, D). In this simple game, checking perfection is easy, since only one mistake is possible. With many strategies, however, there are many more possible mistakes to take into account. Similarly, with many players we may need to worry about who is likely to make a mistake. Note that the starting point of this approach is the set of Nash equilibria of the game. It is assumed that players can calculate them and agree to play one. The goal now is to rule out all those Nash equilibria that are not reasonable agreements. In principle, an equilibrium that is reasonable under a given criterion of reasonableness might cease to be such under another, more restrictive criterion. Specifying why an equilibrium might be unacceptable is made easier by taking into account what would happen if one or more players were to “deviate” from end p.195 the agreed-upon solution. The reason is intuitive: a player should not agree to play his part in an equilibrium if—were the unexpected to happen—he would have been better off by playing another strategy. Therefore we may say that a crucial property required of an
equilibrium is that it is stable to players' deviating from it. To reason about “deviations” from equilibrium, it is helpful to have a richer description of the game. This is the reason why most of the refinement literature refers to games in extensive form, where the order in which players move and the information they have when making a choice are made explicit.
Games in Extensive Form The extensive form of a game specifies the following information: a finite set of players i = 1,n, one of which might be nature (N); the order of moves; the players' choices at each move and what each player knows when she has to choose; the players' payoffs as a function of their moves; finally, moves by nature correspond to probability distributions over exogenous events. The order of play is represented by a game tree T, which is a finite set of partially ordered nodes tT that satisfy a precedence relation denoted by <. A subgame is a collection of branches of a game such that they start from the same node and the branches and the node together form a game tree by itself. In figure 10.7, for example, player 2's decision node as well as her moves form a subgame of the original game. Whereas normal form games are represented by matrices, extensive form games are represented by trees. A matrix description shows the outcomes, represented in terms of players' payoffs, for every possible combination of strategies the players might choose. A tree representation is sequential, because it shows the order in which actions are taken by the players. It is quite natural to think of sequential-move games as being ones in which players choose their strategies one after the other, and of simultaneous-move games as ones in which players choose their strategies at the same time. What is important, however, is not the temporal order of events per se, but whether players know about other players' actions when they have to choose their own. In the normal form representation, players' information about other players' choices is not represented. This is the reason why a normal form game could represent any one of several extensive form games. When the order of play is irrelevant to a game's outcome, then restricting oneself to the normal form is justifiable. When the order of play is relevant, however, the extensive form must be specified. In an extensive form game, the information a player has when she is choosing an action is explicitly represented using information sets, which partition the nodes of the tree. If an information set contains more than one node, the player who has to make a choice at that information set will be uncertain as to which node she is at. Not knowing at which node one is means that the player does not know which action was chosen by the preceding player. If a game contains information sets that are not singletons, the game is one of imperfect information. It may also be the case that a player does not remember what she previously did. In this case, the game is one of imperfect recall. All the games we consider here, however, will be ones of perfect recall, in that players will be assumed to remember what they did and knew previously.
A strategy for player i is a complete plan of action that specifies an action at every node at which it is i's turn to move. Note that a strategy specifies actions even at nodes that will never be reached if that strategy is played. Consider the game in figure 10.7 . It is a finite game of perfect information in which player 1 moves first. If he chooses D at his first node, the game ends and player 1 nets a payoff of 1, whereas player 2 gets 0. But choosing D at the first node is only part of a strategy for player 1. For example, it can be part of a strategy that recommends “play D at your first node, and x at your last node.” Another strategy may instead recommend playing D at his first node, and y at his last decision node. Though it may seem surprising that a strategy specifies actions even at nodes that will not be reached if that strategy is played, we must remember that a strategy is a full contingent plan of action. For example, the strategy Dx recommends playing D at the first node, thus effectively ending the game. It is important, however, to be able to have a plan of action in case D is not played. Player 1 may, after all, make a mistake and, because of 2's response, find himself called to play at his very last node. In that case, having a plan helps. Note that a strategy cannot be changed during the course of the game. Though a player may conjecture about several scenarios of moves and countermoves before playing the game, at the end of deliberation a strategy must be chosen and followed through the game. The game of figure 10.7 has two Nash equilibria in pure strategies: (Dx, d) and (Dy, d). This is easy to verify by looking at figure 10.7, the normal form representation of the game. Is there a way to solve the indeterminacy? Representing the sequential version of the game as one of perfect information (figure 10.7) helps to solve it. Suppose player 1 were to reach his last node. Since he is by assumption rational, he will choose x, which guarantees him a payoff of 4. Knowing (by assumption) that 1 is rational, player 2—if she were to reach her decision node—would play d, since by playing a she would net a lower payoff. Finally, since (by assumption) player 1 knows that 2 is rational and that she knows that 1 is rational, he will choose D at his first decision node. The equilibrium (Dy, d) should therefore be ruled out, since it recommends an irrational move at the end p.197
Figure 10.7 last node. In the normal form, both equilibria survive. The reason is simple: Nash equilibrium does not constrain behavior out of equilibrium. In our example, if 1 plans to choose D and 2 plans to choose d, it does not matter what player 1 would do at his last node, since that node will never be reached. The sequential procedure we have used to conclude that only (Dx, d) is a reasonable solution is known as backward induction (Zermelo 1913 ). In finite games of perfect information with no ties in payoffs, backward induction always identifies a unique equilibrium. The premise of the backward induction argument is that mutual rationality and the structure of the game are common knowledge among the players (CK1 and CK2). It has been argued by Binmore (1987 ), Bicchieri (1989 , 1993 ), and Reny (1992 ) that under certain conditions common knowledge of rationality leads to inconsistencies. For example, if player 2 were to reach her decision node, would she keep thinking that player 1 is rational? How would she explain 1's move? If 1's move is inconsistent with CK2,
player 2 will be unable to predict future play; as a corollary, what constitutes an optimal choice at her node end p.198 remains undefined. As a consequence of the above criticisms, the usual premises of backward induction arguments have come to be questioned (Pettit and Sugden 1989 , Basu 1990 , Bonanno 1991 ). One should distinguish, however, the game theorist's informal justification of equilibrium play from players' reasoning to an equilibrium. A rigorous representation of agents' reasoning involves formally modeling their knowledge and beliefs, and encoding this information into a set of axioms. Players will then be represented as logical reasoners who infer (from a set of axioms and inference rules) a sequence of moves consistent with the axioms. A solution is precisely the sequence of moves players infer from their theory of the game (Bicchieri 1993 ). Part of the puzzlement with backward induction is due to the ascription of the informal backward induction argument to the players. Central to the informal argument is an analysis of out-of-equilibrium play, which is used to justify the choice of a given equilibrium (in this case, the backward induction one). If the backward induction argument is formalized as the players' own theory, it gives rise to an inconsistency when coupled with information that a deviation has occurred. The reason is obvious: any consistent theory that lets players infer an equilibrium becomes inconsistent when combined with a statement to the effect that a deviation from equilibrium occurred. It is only when considering deviations that inconsistencies may arise, but an analysis of deviations need not be part of the players' theory of the game. Such a theory may in fact harmlessly assume that players' rationality is common knowledge: The players would still succeed in computing the backward induction equilibrium (Bicchieri and Antonelli 1995 ). Spurred by this problem, in recent years much foundational work has been devoted to carefully modeling players' knowledge and reasoning in games. Depending on how knowledge and hypothetical knowledge are represented, different authors have reached divergent conclusions about the link between backward induction and common knowledge of rationality. Some of the literature includes Bicchieri 1988b , 1993 ; Aumann 1995 ; Bacharach 1987 ; Stalnaker 1994 ; Samet 1996 ; and Arlo-Costa and Bicchieri 1998 .
Extensive Form Refinements The goal of the refinement program, however, has not been the formalization of players' reasoning. The arguments proposed have been informal, their purpose being the elimination of implausible equilibria. In the normal form, Selten's trembling-hand perfection requires players to check how a strategy will perform were another player to take an action that has zero probability in equilibrium. In end p.199
the extensive form representation, players ask what would happen off-equilibrium, at points in the game tree that will never be reached if the equilibrium is played. In both cases, the starting point is an equilibrium, which is checked for stability against possible deviations. By its nature, the Nash equilibrium concept does not restrict action choices off the equilibrium path, because those choices do not affect the payoff of the player who moves there. For example, the equilibrium (Dy, d) in the game of figure 10.7 lets player 1 make an irrational choice at the last node, since that choice is not going to affect his payoff (which is determined by his choosing D at the beginning of the game). However, the strategy of a player at an off-equilibrium information set can affect what other players choose in equilibrium. Suppose the players consider agreeing to play (Dy, d). In order to choose D, player 1 must decide what would happen were he to play A instead. To decide whether D is a rational move, 1 has to think about player 2's choice at an off-equilibrium node. His conclusion about 2's choice will affect his own choice. But player 2's choice will depend upon how she interprets 1's off-equilibrium move. Player 1, in turn, must be able to anticipate 2's interpretation of his deviating from the equilibrium path. For example, if 2 were to interpret the deviation as a mistake, would she still play her part in the equilibrium (Dy, d), and choose d? If she expects y to be played at the last node, a is a best reply. But, at his last node, why would rational player 1 choose y? Is the agreement to play (Dy, d) reasonable? The earliest refinement proposed to rule out implausible equilibria in extensive games of perfect information is subgame-perfection (Selten 1965 ). A Nash equilibrium is subgame-perfect if its component strategies (when restricted to any subgame) remain a Nash equilibrium of the subgame. The equilibrium (Dy, d) is not subgame-perfect: in the subgame starting at the last node, y is a dominated strategy. Note that the backward induction equilibrium is always subgame-perfect. Subgame perfection, however, applies only (nontrivially) to games that have proper subgames. Any Nash equilibrium of a game without proper subgames is trivially subgame-perfect (since the whole game can be considered a subgame), but in this case the criterion does not help in resolving the indeterminacy. In figure 10.8 , for example, both (c, L) and (a, R) are (trivially) subgameperfect equilibria. The game of figure 10.8 is a case in which we would still like to eliminate equilibria that require players to behave suboptimally in parts of the game that are reached with zero probability if a given equilibrium is played, but cannot be considered subgames. Kreps and Wilson's (1982 ) sequential equilibrium is an answer to this problem. A sequential equilibrium is a combination of strategies and beliefs such that each player has a belief (a probability assessment) over the nodes at each of his information sets. At any information set x where i has to play—given player i's beliefs at x and the equilibrium strategies of the other players—i's strategy for the rest of the game must still maximize his expected payoff. As end p.200
Figure 10.8 players move through the game tree, they rationally update their beliefs using Bayes's rule. The problem with the notion of sequential equilibrium is that—provided beliefs are revised according to Bayes's rule—no further restriction is imposed upon them. The consequence is that far too many Nash equilibria are still considered admissible or “reasonable.” At player 2's information set, if for some reason she assigns a higher probability to node y than y', then her optimal choice is L. If instead she judges y and y' to be equiprobable, she will choose R. Thus both (c, L) and (a, R) survive as sequential equilibria, since each is supported by some acceptable belief. Another common refinement is Selten's perfect equilibrium (Selten 1975 ). In this case players are explicitly assumed to interpret deviations from equilibrium play as “mistakes” and respond accordingly. A perfect equilibrium must be robust to small perturbations of players' equilibrium strategies. Selten's notion, however, by not imposing restrictions upon players' beliefs, lays itself open to the same criticism addressed to Kreps and
Wilson's refinement. If there are several possible “mistakes” a player can make, and beliefs are unrestricted, some equilibria cannot be ruled out simply because they are supported by beliefs that make some mistakes more likely than others. Suppose that in our example player 2 believes that player 1 intends to play c with probability 1–p–q very close to one, but can play b by mistake with higher probability (p = 2/100) than playing a by mistake (q=1/100). If this is what 2 believes, she should choose L. Given her beliefs, L has an expected utility of .06, whereas R has an expected utility of .01. Both equilibria therefore survive some perturbations. It may seem that introducing a restriction as to which mistakes are more likely to occur would help in limiting the number of plausible Nash equilibria. Myerson's proper equilibrium, for example, is a Nash equilibrium that is robust with respect to “plausible” deviations only, by which he meant deviations that do not involve costly mistakes (Myerson 1978 ). Yet what makes a mistake costly is not a simple matter, since it depends upon what players believe other players will do in reaction to a mistake. Consider again the game in figure 10.8. If a deviation from (c, L) were to occur, player 2 would keep playing L only if she believed the probability of mistake b to be greater than that of mistake a. And if player 1 were to believe that 2 would respond to a deviation with L, mistake b would indeed be less costly for him than mistake a. In this case, again, both equilibria will survive. Note, however, that strategy b is strictly dominated by c for player 1, therefore it is highly unlikely that 1 would play b. This example is meant to stress the importance of restricting off-equilibrium beliefs. Such beliefs, too, should be rationally justified. The equilibrium (c, L) must be eliminated because it is supported by the belief that a dominated strategy will be played off-equilibrium, and this belief is inconsistent with common knowledge of rationality. The need to add a condition of plausibility for off-equilibrium beliefs motivated the forward induction refinement (Kohlberg and Mertens 1986 ). Off-equilibrium beliefs, for example, should be consistent with common knowledge of rationality and any inference one may draw from it. A deviation from equilibrium should therefore be interpreted, whenever possible, as a rational move. In our example, player 1's deviation from the equilibrium (c, L) should not be interpreted as a mistake, but rather as a signal that he intends to play a (and get a higher payoff). In this case player 2 would respond with R. The conclusion is that equilibrium (c, L) is not robust to deviations, and should be eliminated. As I mentioned at the outset, the refinement program attempts to establish stability criteria for Nash equilibrium. It presupposes that players will choose to play a Nash equilibrium after having eliminated several alternative equilibria on the ground that they are unreasonable. What counts as a reasonable equilibrium, end p.202 however, depends upon how off-equilibrium behavior is interpreted. This, in turn, hinges on players' out-of-equilibrium beliefs. A formalization of the off-equilibrium deliberation process requires the use of counterfactuals (from the viewpoint of playing a given equilibrium, deviations are contrary-to-fact events). Some work in this direction has been done by Selten and Leopold (1982 ), Bicchieri (1988a ), Harper (1991 ), Samet (1996 ),
Stalnaker (1994 , 2000 ), and Skyrms (1990 , 1998 ). So far we have developed no comprehensive theory of out-of-equilibrium behavior that indicates, for example, when a deviation should be interpreted as a signal and when as a mistake. Such theory would supply substantive (as opposed to merely formal) rationality criteria for players' beliefs, and would thus expand the traditional notion of practical rationality to include an epistemic component. This theoretical inadequacy undermines the eductive goal of inferring (and predicting) equilibrium play from rationality principles alone.
Repeated Games The games we have discussed so far were all one-shot games. Many real-life interactions, however, are ongoing ones. Interesting social phenomena such as cooperation, retribution, trusting and reciprocating, committing and promising can be represented by means of repeated games. Take the Prisoners' Dilemma (see also Dreier, chap. 9, and Sorensen, chap. 14, this volume) depicted in figure 10.1. The label C now stands for “cooperate” and D for “defect.” 1 What should a rational player do? Notice that the payoffs represent ordinal preferences; thus by just looking at the matrix it is clear that each player prefers to defect, whatever the other does. In fact, to defect is a dominant strategy for both. But notice that, if both were to cooperate, their payoffs would be superior to those of joint defection. Mutual cooperation is Pareto-optimal, but it is not an equilibrium. Of course the players may have different preferences, and like cooperation better than defection, but then we would have to draw a different matrix. In the game of figure 10.1, players' preferences doom them to an unsatisfactory outcome. Suppose now that the Prisoners' Dilemma is repeated N-times between the same two players. Since the game is noncooperative, the players cannot enter into binding agreements, so whatever threat or promise they make to each other, it must be enforceable to be credible. Suppose that it is common knowledge among the players that the game is ending at the Nth repetition. In this case, both know that in the last repetition they will both defect, irrespective of what they said or end p.203 did before. At the N–1th stage then, they will both know what will happen at the Nth stage, so the rational choice will be to defect there, too. Proceeding inductively, both players will reach the conclusion that each will (and should) defect at each stage. The players use backward induction (see also Sorensen, chap. 14, this volume) to reach this conclusion, and the only Nash equilibrium of this repeated game is to defect throughout. Since the payoffs to each player are the sum of what she receives at each stage game, the players will end up with 1N each, whereas joint cooperation would have netted them 3N each. I am assuming here that players have common knowledge of the game, payoffs, and mutual rationality. Relaxing any of these assumptions will give us different results, one of which is that players can rationally sustain cooperation for most of the game (Kreps and Wilson 1982 ; Bicchieri 1993 ).
Keeping firm the traditional assumptions, however, we can still achieve cooperation in a repeated Prisoners' Dilemma, provided the game is infinitely repeated or, alternatively, there is an unknown termination and a sufficiently high probability that the game continues another stage. To see why cooperation can be achieved, suppose that, after each repetition, there is a probability of 1/3 that this is the last stage. Then the probability that the game will persist until at least the Nth stage is (2/3) N–1 . Note that the probability that the game is eternal is zero, since (2/3) N → 0 as N → ∞. Now consider the following strategy S: “Play C as long as the other plays C, but if he plays D, play D forever after.” If both players adopt S, they will cooperate forever. Player i's expected payoff of playing C at each stage is 3 + 3(2/3) ++ 3(2/3) N−1 + 3(2/3) N + 3(2/3) N+1 +.Can a player gain by deviating? Suppose player i plays D at the Nth stage. If the other player plays S, the deviant player will at most get the following payoff: 3 + 3(2/3) ++ 3(2/3) N−1 + 4(2/3) N + 1(2/3) N+1 +.For a deviation to be unprofitable, it must be the case that the expected payoff of continuous cooperation (enforced by S) is greater or equal to the expected payoff of, say, deviating at the Nth stage. In our example, C − D = (3–4) (2/3) N + (3–1) (2/3) N+1 + (3–1) (2/3) N+2 +, which amounts to (2/3) N (−1+4). Since this value is greater than zero, this means that the expected payoff of continuous cooperation is better than the expected payoff of deviating at the Nth stage. Therefore (S, S) is a Nash equilibrium. In fact, we know that in our game any feasible expected payoff combination that gives each player more than 1 can be sustained in equilibrium. This result is known as the Folk Theorem, so called because it is uncertain who proved it first but it has become part of the folk literature of game theory. What is rational in a one-shot interaction may not be rational when the game is repeated, provided certain conditions obtain. The conclusion we may draw is that phenomena such as cooperation, reciprocation, honoring commitments, and keeping one's promises can be explained as the outcome of rational, selfend p.204 interested choices. In repeated interactions, even a person who cares only about material, selfish incentives can see that it is in his interest to act in a cooperative way. Whether this is a good, realistic explanation of prosocial behavior is another story. NOTE 1. The most familiar version of the Prisoners' Dilemma is found in Luce and Raiffa 1957 , 94–97. end p.205
chapter 11 PRACTICAL REASONING AND EMOTION Patricia Greenspan The category of emotions covers a disputed territory, but clear examples include fear, anger, joy, pride, sadness, disgust, shame, contempt, and the like. Such states are commonly thought of as antithetical to reason, disorienting and distorting practical thought. However, there is also a sense in which emotions are factors in practical
reasoning, understood broadly as reasoning that issues in action. At the very least emotions can function as “enabling” causes of rational decision making (despite the many cases in which they are disabling) insofar as they direct attention toward certain objects of thought and away from others. They serve to heighten memory and to limit the set of salient practical options to a manageable set, suitable for “quick and dirty” decision making. Current research in neuroscience and other areas indicates that practical reasoning in this sense presupposes normal emotional development and functioning (see, e.g., Damasio 1994 ). Evolutionary accounts of emotions (e.g., Cosmides and Tooby 1992 ) stress their role in the rational design of the human organism. Contemporary philosophy of emotion attempts something stronger, however, in according emotions a role in practical reasoning. Making this an integral role—understanding emotions as functioning within practical reasoning rather than just as spurs to it—means interpreting emotions in normative terms, as providing or expressing potential reasons for action, and as themselves subject to rational assessment and control, contrary to the traditional view of emotions as “passive” phenomena.
1. Emotions as Evaluative The dominant approach in contemporary philosophy rests on assigning emotions an evaluative content (see, e.g., Bedford 1957 , Solomon 1976 , Lyons 1980 , Budd 1985 , Davis 1987 , de Sousa 1987 , Roberts 1988 , Greenspan 1988 , Nussbaum 1993 , Stocker 1996 ). Human emotions typically are directed toward “intentional objects” in the sense of being about something—real or imagined, but in any case an object of thought. Emotions that represent their objects in some positive or negative light (as most do) may be said to have a content expressible by an evaluative proposition. Where the evaluation either is or implies an evaluation of some future contingency that the agent (the one who undergoes the emotion) can bring about or avoid, the emotion provides a reason for or against action. The rational bearing of emotions on action in these terms can be captured in simple form by the first or major premise of an Aristotelian practical syllogism, Aristotle's three-line schema of practical reasoning (Aristotle 1984 : Nicomachean Ethics, bk. 7; for an alternative model, see Thagard 2000 ). In contemporary terms the major premise expresses a pro- or con-attitude toward something that the second or minor premise tells one how to attain or avoid. It evaluates something as good or bad, desirable or worth avoiding—or simply as an object of a current desire or aversion. For instance, fear may be said to represent its object (what one is fearful of or about: a certain event, or a likely cause of it) as a threat, a possible harm. Anger represents its object (another person, or his performance of some action) as already a cause of harm or offense that now calls for retaliation. Despite differences between the propositional content of these two emotions—one describes a possibility, the other a fact (Gordon 1987 )—both include a reason for avoiding some future contingency: the presence or occurrence of what is feared or a failure to retaliate for the cause of anger.
Emotions also involve a corresponding affective element, derived from pleasure or pain on Aristotle's account of emotions (1984 : Rhetoric), that amounts to a good or bad state of the agent and hence supplies a reinforcing reason for action (Greenspan 1988 ). In contemporary decision-theoretical terms, the affective element modifies the “payoff structure” of the situation: the array of potential costs and benefits of alternative responses open to the agent. The discomfort experienced in fear or anger, for instance, provides the agent with a further (pro tanto) reason for acting to change the situation that provokes discomfort. On the other hand, joy or pride, as positive evaluations of some state of affairs or of oneself, do not provide a reason to change anything, but their affective aspect yields a further reason to sustain the conditions that make the evaluation appropriate. However, the focus of the agent's attention in all these cases is normally the evaluative content of emotion, not her own state of feeling. end p.207 Apart from its role as a source of reinforcing reasons, the affective or feeling aspect of emotion is useful to practical reasoning just insofar as it serves to hold in mind the evaluative content of emotion without explicit reflection. For instance, while driving on the highway one does not have to deliberate at length about possible bad consequences of swerving and their relevance to the task of steering straight. To the extent that steering straight is not just automatic, and the driver wanders out of her lane, an anxious awareness of the possibility of an accident brings her quickly back to the task. This way of anticipating practical eventualities in everyday life corresponds roughly to neuroanatomist Antonio Damasio's (1994 ) understanding of emotions as “somatic markers.” Damasio's view is put forth to explain why cases of emotional impairment due to brain lesion (such as the famous nineteenth-century case of Phineas Gage) involve a loss of practical reasoning ability. Emotions serve to “mark” practically significant thoughts with bodily (and hence affective) indicators of past experience. According to an evaluative account, characteristic thoughts have come to be contents of emotion—and part of what identifies them as the types of emotion they are: fear, anger, joy, pride, and so forth. A question for philosophers who accept an evaluative account is whether and in what sense emotions themselves are subject to rational assessment. It is only in a qualified sense that an irrational evaluative stance could be said to justify action. Hume's famous denial (see 1978, bk. 2, pt. 2, sec. 3) that the passions can be rationally assessed at all was based on his interpretation of them as nonrepresentational. But contemporary evaluative accounts interpret emotions, in effect, as representing evaluative propositions: that something portends harm or deserves retaliation or is a stroke of good luck or praiseworthy.
2. Belief-Based Views The evaluative implications of emotion have been characterized thus far without commitment to a particular contender for capturing their essential nature. This is the
philosopher's usual first question, and it was set up as the focus of contemporary debate on emotion by James (1884 ). However, as Rorty (1980c ) and others have noted (cf. Griffiths 1997 ), emotions do not form a natural class. A later approach (de Sousa 1987 , Greenspan 1988 ) begins with questions of rationality, treating cases of rational emotional response as paradigmatic. Emotions in lower animals or infants (along with some elements of our own experience that end p.208 derive from them unchanged) can be understood from this standpoint as deficient instances of full-blown emotions functioning normally in adult human life. As applied to accounts of the nature of emotions, the evaluative approach is often discussed under the heading of “cognitivism,” which interprets emotions themselves as amounting to or containing cognitions, usually seen as mental states representing evaluative propositions. The most straightforward version of the cognitivist view is “judgmentalism,” which understands the nature of emotions in terms of judgments: assertive propositional attitudes, assessable by the ordinary evidential standards applied to beliefs. The view traces back to the Stoics (see Nussbaum 1993 ), though the Stoics interpreted our ordinary emotions as “confused” rather than rational judgments. The contemporary version of judgmentalism was introduced into the Anglo-American philosophical literature by Bedford (1957 ), in a Wittgensteinian response to the Cartesian tradition of identifying emotions with particular passions or feelings, as in Hume (cf. Descartes 1984a , James 1884 ). The view was extended by Kenny (1963 ) and Pitcher (1965 ), and it was brought into connection with other philosophical traditions and with work in psychology and other areas by Solomon (1976 ), Neu (1977 ), Lyons (1980 ), and Ben-Ze'ev (2000 ). Davidson (1980 ) defends a version of judgmentalism for pride and related emotions. Proponents of judgmentalist accounts do not always seem to have the same thing in mind by “judgment.” Diverse cases indicate, moreover, that the view needs some sort of qualification to capture situations where emotions register evaluative thoughts that may not rest on evidence supporting judgments in the sense of all-things-considered beliefs but nonetheless seem to be rational to hold in mind and to act in light of. Some irrational cases such as phobias (fear of a dog one knows to be harmless, say) that undermine judgmentalism as a view of the nature of emotions are explored at length in Pitcher 1965 , Stocker 1987 , and Greenspan 1981 and 1988 . However, there are also many morally or motivationally important emotions that one might be inclined to assess as rational but that turn on imagining some state of affairs that is not applicable (at least in current terms) to oneself. Cases of empathetic emotion, for instance, seem to involve putting oneself in the position of another person, typically someone suffering harm. However, acting effectively out of sorrow felt on behalf of someone else often depends on not losing sight of the fact that one is not really in the same position oneself. Similarly for anticipatory emotions such as guilt at the thought of something one might (or might not) do. There are also “emotional self-management” cases—of generating upbeat feelings, say, by “looking at the bright side” of some more problematic object of evaluation, or looking past someone's flaws to sustain love—that are crucial to mental health and to thriving interpersonal relationships.
end p.209
3. Modifying Judgmentalism for Rationality Issues One might distinguish a possible variant of judgmentalism meant to apply specifically to rationality issues. Whatever their nature as psychological states, according to this account, emotions can be understood in terms of some sort of claim they make about the agent's situation. Whether a given emotion is rationally appropriate depends on whether this corresponding judgment would be warranted as a belief—whether or not the emotion actually involves that belief. Emotions that do not conform to this model, such as empathetic or anticipatory emotions, might still sometimes be said to play a limited instrumental role in practical reasoning as cases in which an irrational mental state serves as part of an overall rational strategy. This idea of emotional motivation as “rational irrationality” comes up particularly in discussion of evolutionary design strategies that might account for emotional proclivities underlying human morality (see Frank 1988 ). Thus, one could grant that feeling sympathetic sorrow over someone else's loss would not entail a belief that one has suffered some sort of loss oneself, or warrant for such a belief. But on the present suggestion that is just because it is not really appropriate to one's own situation. The feeling may still be important to sustaining mutually supportive personal relationships, and we can encourage it in ourselves with that end in mind, though to preserve its other-directed quality, we usually try to keep such calculations subliminal, or refer them to evolutionary design. Whether this is a satisfying answer for all cases depends in part on how we think of rationality. In common parlance the term is sometimes used in a “thick” sense that would be implausible for most emotions insofar as it connotes explicit calculation; by contrast, Gibbard 1991 uses the term in an extremely “thin” sense in which it conveys just general positive endorsem*nt. But there seems to be room between these poles for a use that captures distinctions we commonly make between reasonable and unreasonable emotional reactions, appropriate and inappropriate in what seems to be roughly a rational sense, having to do with some notion of fit to the circ*mstances that constitute grounds or evidence for emotion. We might think of this medium-thick sense as “representational” rationality insofar as it depends on assigning emotions intentional content. However, our ordinary notion of emotional appropriateness seems to allow for variable standards of evidence and for more than one option in many cases. For instance, under ordinary circ*mstances it might be reasonable enough to feel somewhat worried about how a recent medical test will turn out, even in the absence of grounds for real agitation. At the same time, there may be no rational end p.210 barrier to foregoing any thought about the matter or even feeling hopeful or confident. Being seized with fear would be irrational, let us say, not just because it would do no
good, but more fundamentally because there are no grounds for holding in mind with that intensity any subset of the available evidence that would warrant the belief that one is in danger. An emotion might come out as rationally appropriate, according to this way of looking at things, even in cases where the corresponding all-things-considered belief would not be warranted. In other words, emotional appropriateness does not seem to be determined by a unitary overall assessment of the available evidence, of the sort presupposed by belief warrant. There is at most one adequately warranted belief (over the threshold for assent) about the probability of a bad outcome, relative to the evidence available to the agent. Belief may admit degrees, but where the evidence is slight, it would be irrational to adopt even a low-grade belief, if that entails adopting conflicting beliefs. However, the emotional analogue (ambivalent feelings, in some cases even intense feelings) would seem to be rationally acceptable—rational in a relatively undemanding sense that allows for options (Greenspan 1980 ). We can think of this as “perspectival” rationality. Moderate worry and hope in the case just outlined may each be rational from different perspectives—with reference to different subsets of the available evidence, each worthy of the relevant degree of emotional attention, though not requiring it. A perspectival account of emotional rationality would appeal to a notion of warrant for a prima facie belief, as determined from some practically significant standpoint of evaluation (Greenspan 1988 and 1995 ). So it still represents a fairly conservative move away from judgmentalism on rationality issues—though it does move beyond the view just distinguished, which makes out emotional appropriateness as dependent on warrant for a corresponding belief. The perspectival account associates emotions with evaluative thoughts rather than strictly judgments in the sense of beliefs (or acts of forming beliefs), but it still allows that the rationality of emotions can be explained in terms of belief warrant. 4. A Causal/Historical Approach In a more radical departure from judgmentalism, emotions might be thought to be justified rationally by reference to their causal histories rather than some notion of evidence of the sort that applies to beliefs. In light of the prominence of causal theories in other areas of philosophy, a causal/historical approach to emotions would seem to provide an alternative to judgmentalism and its variants. Someapproach along these lines is suggested by recent evolutionary treatments of emotion in philosophy (e.g., Gibbard 1991, Griffiths 1997 ). However, the only example of it that has been formulated to apply to issues of rationality in the sense indicated is the “paradigm scenarios” view in de Sousa (1987 ), which is still phrased in roughly cognitivist terms. According to de Sousa's account, whose initial sources are psychoanalytical, emotions are based on infantile response tendencies such as smiling or crying that become intentional or object-directed and thereby come to be signs or elements of emotions by association with a typical story about what prompts the response. The story is what de Sousa calls a “paradigm scenario,” meaning that it sets the standard for a particular emotion type, essentially by establishing criteria of appropriateness. Originally, what is involved here is the child's interaction with caretakers, which serves to modify natural responses by a form of dramatic role-playing (eliciting or exchanging smiles, say, or conveying empathy). Later cultural influences can modify the emotion, though it will remain appropriate in some form as a response to the paradigm scenario.
A central example de Sousa wants to defend as appropriate in these terms, in response to some feminist arguments, is jealousy, or jealous rage, with psychoanalytic sources shaped by monogamous culture. According to his account the emotion is set up as a response to a situation of “being robbed by another of vital physical attention” (185). That is how it gets its status and meaning as an intentional response, so it necessarily counts as appropriate on that basis. It is unclear, though, whether the paradigm scenarios account can explain more than the appropriateness of (at least some central cases of) a given type of emotion. De Sousa's essential point about rationality in the sense of appropriateness seems to be just that there cannot be an emotion type of which all possible instances are inappropriate. However, it does not follow from this that a given emotion instance should be assessed as appropriate on the basis of its resemblance to the paradigm scenario for its type. More specific rationality questions—under what circ*mstances jealousy is appropriate in adult life, say—remain open. De Sousa speaks of the rationality of emotions in several senses, but overall he seems to be most concerned with rationality in the sense of objectivity. On that interpretation, emotions are rational in general terms to the extent that they represent real properties of the world, existing independently of emotions. This amounts to an emotion-based version of the metaethical question of the objectivity of values, since the properties in question, corresponding to different types of emotion, are “axiological” or value properties. However, recent views within the metaethical literature would seem to allow for emotional rationality in a sense independent of objectivism about values (see, e.g., Gibbard 1990 , Blackburn 1998 ; cf. Greenspan 1995 and 1998 , D'Arms and Jacobson 2000 ). Objectivity need not be our concern when we speak of the rationality of particular emotions in the sense of appropriateness, but rather their end p.212 fit to the reasons available to the agent, leaving it as a further question whether reasons or values in general can be thought of as objective. De Sousa at some points limits himself to a claim of “minimal rationality” or basic intelligibility of emotion types, but he also applies the paradigm scenarios view to the question of the correctness or incorrectness of particular emotion instances, which his discussion makes out as dependent on the notion of normality (cf., e.g., 187, 201). But some emotional responses that are normal and understandable might be inappropriate. We can make sense of a claim, for instance, that the paradigm for male jealousy—what degree and sort of attention the jealous agent feels he is entitled to and hence can be “robbed” of—has been distorted by male upbringing in a certain culture. So even the received paradigm of an emotion can be skewed or off the mark. Applying the paradigm scenarios approach to cases would seem to involve appeal to an interpretation of the situation that might be summed up in a proposition: that one has been robbed of vital physical attention, in the jealousy example. How this gets extended or modified by culture (extended to other forms of attention, perhaps, or modified by challenging the criteria of entitlement to attention) may be a matter of historical derivation, but the criteria at a given time and relative to a given culture would seem to be specifiable in terms that fit the belief-based approach outlined earlier. De Sousa rejects
the analogy between emotions and beliefs as a basis for his approach to rationality issues at least partly because he understands rational beliefs as arrived at in a certain way, on the basis of reasoning (see 5; cf. 197ff.). However, this is not part of the belief-based approach as here intended. It is not clear how the paradigm scenarios view could handle cases without appeal to belief-based criteria of rationality. For example, imagine someone who momentarily feels jealous anger when his wife exchanges glances with another male at a party. In de Sousa's terms the man is reacting to what he sees as a robbery of vital attention—in terms of Aristotle's definition of anger (see 1984 : Rhetoric, 378a31–b5), an unjustified slight, or at any rate an indication that a slight is imminent. But this is an evaluative judgment interpreting past events and their natural and conventional meanings: what a glance means or can mean, what legitimate expectations a relationship confers, that a glance involves or might lead to intimacies that violate those expectations, and that violating them inflicts loss and insult on the agent. The emotional reaction in such a case might sometimes be assessed as inappropriate to the current situation—if the agent really knows, say, that his wife and the recipient of her glance, a colleague in her area, are merely reacting to a professional faux pas on the part of someone else at the party—even where situational cues do naturally give rise to jealousy because of their superficial resemblance to the paradigm scenario. To explain which cues render an instance of emotion appropriate, rather than merely natural or normal or understandable, we end p.213 seem to need at least implicit reference to the notion of a propositional content, as what the emotion still essentially “claims” about the situation, even if the reasons for it have changed since the paradigm scenario was established.
5. Emotional Strategies A variant of the general causal-historical approach that rejects reference to propositional attitudes is the evolutionary account in Griffiths 1997 . Though Griffiths is mainly concerned with the question of the nature of emotions rather than rationality issues, his argument turns out to rest at crucial points on a familiar way of setting up emotions in opposition to reason. He continually adverts to a definition of emotions as “irruptive” or passive states interfering with the agent's stable goals and not themselves subject to rational strategy (see, e.g., 155–57, 233ff., 242ff.; cf. 9, 16, 118, 120). Those cases in which emotions appear to be used strategically cannot be genuine emotions according to his account. Griffiths particularly has in mind emotions based on the social pretense that we lack rational control over a certain pattern of behavior. He draws on anthropological examples from defenders of “social constructivist” views of emotions (see, e.g., Averill 1980 , Lutz 1986 ) in which the conditions of particular syndromes of emotions seem to be based on social conventions. A parallel from contemporary urban life is the “Rambo” syndrome of
violent male rage; some conceptions of romantic love might also qualify. However, Griffiths's point is that whatever is felt in such cases would be disqualified from inclusion in a class of genuine emotions. However, it seems, on the contrary, that we can make sense of a self-fulfilling social pretense of emotion in the example of male jealousy drawn from de Sousa. Imagine someone working himself into states of jealous anger on flimsy or imagined grounds (or even on good grounds that he normally would ignore) in order to provoke a certain kind of interaction with his spouse—to exert control, perhaps, or perhaps just to bring about an occasion to express and enhance affection. An attempt to understand emotions as voluntary actions, things we do rather than states that come over us, was part of Solomon's (1973 ) version of judgmentalism, based on Sartre's (1962 ) conception of emotions as “magical transformations of reality” but meant to allow for rationality. However, we need a less extreme version of the view to modify the usual picture of rationality and emotions. The rational role of emotions depends on the fact that our control over them is limited. Thus, for instance, Frank (1988 ) assigns emotions a crucial role in answer to end p.214 what he calls “the commitment problem,” the problem of how we can demonstrate to others effectively our commitment to future actions against narrow self-interest, as needed to secure their willingness to take part in cooperative schemes (on the model of the Prisoners' Dilemma, where each agent has to resist inducements to confess a joint crime). Emotions are as useful as they are to social communication just by virtue of the fact that they are not states an agent can simply produce at will. So the question is whether we can strategically bring about genuine states of partial uncontrol, individually or collectively. But the answer seems to be that we do this all the time. I would add to Frank's account the suggestion that we can do it without sacrificing rationality—even in the short term, to the extent that we undergo an emotion. The point depends on understanding emotions themselves as potentially rational and as rationalizing action, but also on appreciating features of their role in directing attention that make them subject to manipulation by the agent (cf. Mele 1995 ). An effective emotional strategy has to be somewhat indirect: we cannot just “talk ourselves into” sincere emotional states by commanding ourselves to feel them or by taking immediate aim at the signs of emotion. Instead, we set up the conditions under which they would arise. These include conditions of thought as ways of understanding the situation. Imagine working up an emotion in order to argue effectively. A case I used in Greenspan 2000 involves letting oneself get angry about a consumer complaint when the time comes to confront someone at the store. But one might also consider the everyday occasions in teaching or giving a paper when one works up a strategic emotion: indignation at opposing views or enthusiasm for a question that is not of current concern. This might be (and remain) just a pretense, but it also might be a real result of focusing attention on aspects of the material and the surrounding situation that are likely to generate the requisite emotion: the wrongheadedness of the opposing view or the features of one's own view that initially made it seem worth defending.
Some people are better at this than others, needless to say. Part of what it takes is control over attention—not entertaining those self-defeating thoughts one may well have on hand, of sympathy for the opposing view and despair at the imperfections of one's own. Also, it is crucial not to focus too directly on strategy: the aim of generating an emotion in oneself that will enhance performance. To count this restriction of attention as rational in decision-theoretical terms, we have to be clear that rationality does not require being immediately and centrally aware of everything one knows. According to standard assumptions in decision theory it does require full knowledge pertaining to the choice situation. But putting blinders on oneself in this sense (or passively accepting them) is “within reason,” as long as what one knows remains within reach, perhaps on the periphery of attention. end p.215 A perspectival evaluative account of emotions and emotional rationality can accommodate such maneuvers. We have two competing partial views of the situation— one focusing on positive features, one on negative—and some choice about where to focus. So for different purposes we can shift the focus, thereby generating the relevant emotional state—not with any certainty but with enough probability for the maneuver to count as an exercise of rational control. Since the state involves positive or negative affect (or possibly a mixture of both), we have essentially modified the payoff structure by adding emotion to the preexisting situation. This account is compatible with the usefulness of emotions as commitment devices, which requires that our control over them be incomplete. For purposes of commitment it is enough that emotions be difficult to fake convincingly and that they make certain courses of action (or refraining from action) difficult. This means that emotional states, or those that are useful as commitment devices, have to exhibit a certain momentum: we cannot so easily just talk ourselves out of them, including those we managed to talk ourselves into. In the case of generating enthusiasm to give a paper, momentum (with any luck) will be a product of the rewarding aspect of the feeling, as augmented by others' positive responses. But in a commitment case like Prisoners' Dilemma something more is possible: if the prisoners have sworn loyalty to each other and firmed up their feelings accordingly, that means instilling negative reactions to the thought of betrayal that will be hard to go against when the opportunity to gain by confessing arises. This case is in some ways easier in that it involves inducing dispositions to feel certain emotions rather than occurrent emotions. The notion of cultivating certain traits in ourselves, including emotional dispositions, is familiar from Aristotle's account of habituation in virtue (1984 : Nicomachean Ethics) and still plays a role in philosophers' views of emotional motivation (see, e.g., Wollheim 1991 ). However, most recent approaches to emotion focus on occurrent states, perhaps in reaction to an earlier overemphasis on dispositions in philosophical behaviorism (see Ryle 1949 ). To appreciate the rational role of emotion, we need to consider uses of occurrent emotions such as anger (to back up a threat to call the Better Business Bureau, in my consumer case), as well as the sorts of commitments an agent can make to himself—to follow through on a resolution (possibly by “holding onto” his feelings of anger).
Damasio's point about “marking” thoughts with emotion covers some cases of emotional self-commitment. Consider learning a foreign language: one way of implanting information in long-term memory is to focus emotional attention on it—perhaps by thinking of some association that evokes amusem*nt. Examples involving practical reasoning include viewing the prospect of going back on one's resolutions—analogous to betrayal in Prisoners' Dilemma—with self-disgust or some other version of anticipatory guilt, bringing on a motivating emotion now as well as generating a disposition to feel it later.
6. Emotions as Reasons By inducing an occurrent emotion, I would say, we give ourselves a further reason for action—much as we do when we make a promise to others. We can count an emotion as itself a reason in the sense of a motivating reason, a state of thought, or as a source of reasons in a more objective sense, of normative reasons as facts about what needs to be done to sustain or ameliorate a state of emotional comfort or discomfort. However, this rational role of emotions may seem hard to accommodate within the traditional division of views within modern moral philosophy. On the one hand, though it carries echoes of Kantian notion of making law for oneself, it runs very much against the usual picture of Kant's emotionless notion of practical reason (see esp. Kant 1981 , 99; but cf. Sherman 1990 ). On the other hand, while the contrasting Humean picture accords emotions a necessary practical role, this seems to depend on undermining the role of practical reason. However, here as elsewhere an alternative model is provided by Aristotle (1984 : Nicomachean Ethics). Among other things, Aristotelian ethics suggests a picture of practical reasoning that includes not just the calculation of means to ends but also the determination of constituents of valued wholes. Aristotle had in mind happiness, of course, but one might fill in something shorter-term for particular emotions. Action expressive of an emotion has a certain value for the agent as the completion of a more integrated psychological whole: a fit between inner and outer states of oneself, as the original condition from which adult reactions developed. For that matter, action on emotion counts as an instance of value-expressive behavior to the extent that emotions are felt evaluations: they amount to affective states assessable from the standpoint of the agent but also with evaluative thought content. So there are two layers of evaluation that we can appeal to here. My suggestion is not that either of them is final. Expressing one's feelings is certainly not always the most rational thing to do. For one thing, the evaluations that emotions register need not represent one's considered view of the situation. But there is always pro tanto reason for expressing emotion, in the way thinking that p is a pro tanto reason for saying “p,” even if it is ultimately defeated by countervailing reasons. That it manifests a state of mind is a consideration in its favor. end p.217
The reasons that emotions provide according to this account are supposed to be justifying reasons, reasons in a normative sense—good reasons, at least as far as they go. There is some resistance to according emotions this role, in part because it seems to cross categories of emotion versus reason, but also because of its apparent implications for moral motivation. Motivation by moral emotion, if understood in the terms I have suggested, can seem suspiciously self-regarding—as if the agent's goal in acting morally were either to improve her state of comfort or just to express herself. However, if the content of emotion is a moral evaluation or an evaluation that has moral worth (such as one that recognizes the value of another person), I think this objection can be met, at least for any reasonably broad understanding of moral motivation. One might want to say that emotions themselves can at best just heighten the force of nonemotional reasons, whatever reasons there are for the evaluative content of emotion, or for holding it in mind. But consider how undertaking an obligation by promising or the like typically augments such reasons as already may have been in force for acting in the way that the obligation prescribes. Affect adds a further factor for an agent to take into account, a psychological factor under limited control. Within moral philosophy, attention to emotions as subjects of moral judgment has generally emphasized their role as elements of personal moral worth (see, e.g., Oakley 1992), on the assumption that their uncontrolled aspect makes them unfit subjects of moral requirement. Indeed, emotions are often invoked in explaining weakness of will. However, insofar as they involve more or less immediate comfort or discomfort, they can serve to modify the payoff structure of the situation prompting weakness of will by reversing temporal discounting (see Frank 1988 , Greenspan 1988 ; cf. Elster 1999 ). Their uncontrolled aspect can itself serve as a check against weakness of the sort that involves recalculating one's reasons when the time for action arrives. Emotion provides a reason for action that is itself somewhat recalcitrant to reason, though subject to indirect control via control over attention. The idea of emotions as factors in practical reasoning might seem to set up a dilemma. On the one hand, if their rational role is analogous to that of belief, then they are reasons only in a subjective sense, not as facts about the world (including ourselves) but as purported representations of facts. My feeling in the consumer case that the store has treated me unfairly would seem to justify threatening them with a complaint only on the assumption that it reflects the truth of the matter. On the other hand, to the extent that we think of emotions as providing new facts that we need to consider in practical reasoning—my angry feelings in the consumer case as unpleasant feelings I ought to act to assuage—they seem to play an obvious but trivial role, in the manner of “brute” facts like the fact that one has a headache, viewed as a reason (say) for staying home from a disco. end p.218 However, consider what it is like to hate the disco scene and to feel repugnance at the thought of being part of it. The evaluative and affective aspects of emotion (or the facts they represent or constitute) reinforce each other as part of a single reason for declining an invitation to a disco: it would be awful, hence I could not stand it, hence it would be all the more awful—and so on. This combined reason is not subject to either side of the
alleged dilemma, and it provides a firmer barrier to persuasion than either evaluation or affect alone could hope to do. Philosophers who accept a desire/belief account of intentional action—a contemporary reading of Aristotle's practical syllogism that is often thought of as Humean—sometimes interpret the motivational role of emotions in terms of desire (see, e.g., Marks 1982 , Searle 1983 , pp. 31–33, Gordon 1987 , Wollheim 1991 , Green 1992 ). Emotions of course can give rise to desires—a desire to avoid discos, in the case just outlined. In the contemporary philosopher's sense of desires as simple “wants,” which leaves out the element of emotional affect, there is a trivial sense in which this would have to be true of any emotions that result in action. However, emotions do not simply reduce to desires, or desire/belief pairs; we need to bring in affect to explain their full rational bearing on action. Nor need the reasons for emotions depend on the agent's (prior) desires: there may be reasons for feeling something—a moral requirement of compassion toward others or guilt at wrong action—that apply to us regardless of our desires. It may concern some moral philosophers that the bearing of emotions on action according to this sort of account would depend on our actually feeling them. Whether we do or not is of course a contingent matter and not subject to complete control. However, presumably the general moral requirement that is satisfied by feeling a given emotion would also require (perhaps as a second-best alternative) that one muster whatever other motivational resources one can to perform the kind of act that typically results from that emotion: aiding those in distress, say, for compassion, or making up for past wrongs, in the case of guilt. It is important not to confuse the moral or other practical reason to feel a given emotion—whether as a reason for action or just in itself, as part of a personal ideal of virtue—with the reasons why it is appropriate to feel. According to the perspectival account outlined here, the latter are analogous to evidence for rational belief, except that we do not assume it is irrational not to feel an appropriate emotion. At this stage, if rational warrant is all that is in question, we have something like a theoretical syllogism yielding sufficient but noncompelling reason for an emotion. Its conclusion serves, or can serve, as the first premise of a practical syllogism leading to action. end p.219
7. Evaluative Affects My use of the propositional attitudes approach to emotions has been directed in the first instance toward questions of rationality. This is usually taken as a later question, but I think that an answer to it can be read back onto the philosopher's traditional first question, about the nature of emotions, for an answer that does not pull quite so sharply away as judgmentalism did from the traditional approach to emotions in terms of feelings. For recent attempts to reestablish the feeling theory, see Leighton 1985 , Robinson 1995 , and Pugmire 1998 . However, emotions can be feelings and still be propositional attitudes. The propositional attitudes approach need not be understood as assigning to emotions a separable element of propositional thought as distinct from feeling. My
suggestion is rather that emotional affect itself renders a judgment that can be stated in propositional terms—by us theorists, that is, rather than necessarily by the agent. The view depends on assigning the affective element of emotion an intentional content. The assumption of intentionality at this level of basic feeling can sound mysterious, but in principle it is no more so than in familiar cases involving units of language and thought. Insofar as emotional comfort and discomfort does have a content, moreover, the affective element of emotion should not be equated with simple hedonic states of pleasure or pain. Instead of thinking of “content” on the metaphor of containment in a composite entity, the idea of emotional affect as pointing toward something outside itself (though internal to the composite entity, of affect plus evaluation, that we think of as the emotion) might be less misleading, with bodily gestures like pointing as the appropriate analogue. The appropriate cognitive analogue, rather than linguistic meaning, would seem to be reference—as in the thesis from Brentano 1973 that made intentionality an issue, the “intentional reference of the mental.” We need not assume that the relevant referential relationship is type-to-type—that a given sort of affect always refers to the same proposition—as cases of demonstrative or other indexical reference make clear. For emotions like love and hatred whose surface structure is not propositional (one need not love the fact that the love object has such-and-such a feature), we can distinguish propositional components of emotion corresponding to its motivational role: discomfort that one is far from the love object, say, corresponding to the desire for closeness characteristic of love (Greenspan 1988 ). My own rather minimal assumption has been that the positive or negative aspect of a given feeling—understood in terms of motivational significance rather than hedonic tone—represents the positive or negative aspect of the corresponding evaluation. However, there are ways of further specifying this simple bipolar account, which was set up just for purposes of getting at the rational role of emotions. end p.220 The evidence cited by Griffiths (1997 ) from Ekman 1992 —and ultimately from Darwin 1965 —groups our basic emotions into multiple “affect programs” that evolved initially to prepare the body for different modes of action: fight or flight, for instance, in the familiar cases of anger and fear. The element of feeling in these clusters of responses comes to be a sign of, and in that sense “about,” the need for a certain kind of action. This natural referent of feeling would then be subject to social modification, including the sort of individual interaction with caretakers that de Sousa's account of paradigm scenarios takes as explaining the intentionality of full-fledged emotions. This account might be extended to cover purely expressive action, without evolutionary function, possibly mediated by symbols based on social learning. A more mentalistic variant might include or substitute reference to an associated mental act such as directing attention—or simply the act of taking some considerations as reasons for action (cf. Scanlon 2002 ). At any rate, given the broad use of “judgment” in the emotions literature, the general suggestion that feeling itself can have an evaluative intentional content may be thought of as a friendly amendment to judgmentalism. There are parallels to it in the continental literature (cf. Brentano 1973 , 99, 223), though with emotions presented in
contrast with judgments. In general, the view that “affect evaluates” (as I would sum it up in slogan form) affords a way of accommodating empirical findings (as in Zajonc 1980, LeDoux 1996 ) that some cases of emotion, such as primitive fear responses, do not involve cognition (cf. Deigh 1994 , Robinson 1995 ). Cases of “gut feelings” recording good reasons for action suggest that even inarticulate emotions sometimes act in aid of rationality.
NOTE For comments on earlier versions of this paper, I owe thanks to Erich Diese, Scott James, Stephen Leighton, and Jerrold Levinson.
chapter 12 THE RATIONALITY OF BEING GUIDED BY RULES Edward F. McClennen Intuitively, to be guided in one's choice of action by a rule is not simply to choose in conformity with that rule; it is also to choose to engage in or refrain from a course of action for the reason that it is required by, or prohibited by, a rule. In such cases the rule is thought to determine whether a given choice of action on our part is justifiable or acceptable. Of course, there is a substantial issue that needs to be resolved as to what it means to say that there exists a rule that functions in this way. In the case of the rules of a free, liberal, democratic society, however, it is customary to suppose that the rules that function in this way are those that one has accepted or to which one has made a commitment and perhaps even helped to formulate, directly or indirectly. In what is to follow I shall limit myself to a discussion of the issue of the rationality of being guided by rules to which one has made such a commitment. 1 It would seem clear, of course, that not just any kind of rule could plausibly function in this manner and thus provide reasons for the choice of an action. As John Rawls (1955 ) argues in “Two Concepts of Rules,” what are sometimes called “maxims” or “rules of thumb” are not ordinarily thought to provide this sort of reason for action. Rather, they serve merely to summarize past findings concerning the application of some general choice-supporting consideration to a partic end p.222 ular case. Such a rule, he argues, presupposes choice-supporting considerations and cases that are logically prior to, and thus can be described without reference to, the rule. Moreover, the person faced with choosing is always entitled to reconsider the correctness of what the maxim recommends or urges—by direct appeal to the underlying choicesupporting considerations. Such rules, then, appear to serve a purely heuristic function. But there are, on Rawls's account, rules that do provide reasons. One class consists of the rules of practices—social, commercial, legal, and political. Such rules are, he suggests, logically prior to the cases to which they are to apply, and those who participate in such
practices are not at liberty to decide for themselves on the propriety of following the rule in particular cases. Specifically, while the rule itself may be defended by appeal to various considerations—for example, various general, utilitarian considerations—those who are bound by the rule cannot take exception to it by direct appeal to those considerations. That is, one can explicitly invoke a “two-tiered” theory of justification for at least certain practices. One appeals directly to a rule to justify or defend the choice of an action, and while the rule in question is itself subject to a defense in terms of various supporting considerations, one cannot directly appeal to those considerations in order to justify departing from the rule. Rules of this sort are, for example, involved in various sports and recreational games. By reviewing the history of the game of baseball, for example, one can, no doubt, offer some sort of explanation as to why the “three-strikes-and-you-are-out-rule” was incorporated into the game. Whatever those considerations were, you could not stand at the plate and start arguing with the umpire who has just called you out by that rule, urging that consideration of those underlying reasons supported your proposal that this time at bat you be allowed four strikes. Again, one may place a special, intrinsic value on accepting and being guided by certain rules—as in the case of dietary rules associated with a particular religion. But what are we to say regarding the appeal to a two-tiered theory when it comes to rules the following of which is presumed to be of purely instrumental value, that is, productive of various kinds of benefits? The rationality of subsidiary rules within a utilitarian moral and/or political theory is a case in point. But, more generally, it is customary to defend various rules by appeal to what would benefit the individuals within a group who are trying to coordinate their actions in a manner that will prove mutually beneficial. In these kinds of cases we appear to face a fundamental dilemma. Consider any rule R and suppose that R requires an agent X to do A in circ*mstances C. When C obtains, either there is a better option than A available or there isn't. If there were a better option, then it would seem to be irrational to perform the inferior option A. On the other hand, if there is no better option than doing A, that is, if doing A is supported by the balance of reasons, then X is justified in choosing A—not because A is required by R, but because rational end p.223 agents should always pick options that have no better rivals. In short, rule-guided behavior cannot be justified: either the rule gives the wrong result, in which case it is irrational to follow it, or it gives the right result, in which case guidance by the rule is irrelevant. 2 This dilemma presents a challenge to the idea of rule-guided choice within the framework of a theory of rational deliberation in which personal, moral, social, or legal rules are regarded as having an instrumental value. It suggests, in effect, that it is odd for a rational agent to cite a rule as a conclusive reason for choice of an action and then acknowledge that the rule was heeded because of the benefits that rule-following brings. In the absence of attaching a special value to rule-following as such, how can the rule in this case play a role in justifying action? The dilemma is very robust, moreover, because it seems to turn
on nothing more than a very weak assumption, namely that it is irrational to choose an option that is judged inferior to another identifiable and available option. To the dilemma as formulated there would appear to be three distinct responses. First, one can accept the conclusion that rules have no real standing in a theory of the justification of choices. That is, one treats the dilemma as a valid reductio and accepts both of the disjuncts above: rule-guidance is either irrational or redundant. This amounts to adopting an “error” theory of rules, according to which rational rule-guided choice of action is only apparent. 3 All in all, this seems to be an option of last resort, to which one would retreat only if one is convinced that all other possibilities have been exhausted. Second, one can adopt a compatibilist strategy and grant that rational agents should always perform the action supported by the best reasons but insist that rule-guidance can be very relevant to instrumental practical reasoning. The compatibilist assumes that every choice that the agent can make is associated with reasons for and against. These reasons can be compared and, most importantly, aggregated. The choice that is justified, then, is the one that is supported by the greatest balance of pro over con reasons. On this view, rules can provide pro reasons for acting in accordance with them by appeal to a variety of practical considerations, chief among which is that they can help us to coordinate our actions over time, that is, execute in a coherent fashion the plans we adopt, coordinate our actions with others, and save on deliberation costs. The compatibilist, then, denies that if the rule gives the right result, then the rule is irrelevant or redundant. Thus the compatibilist strategy allows one to preserve the idea that rational agents should always be faithful to the underlying reasons for action, while also providing a role for rules. The compatibilist is still committed to the view, however, that if allowing some rule to guide our choice was not supported in a given case by the balance of reasons, then it would be irrational to be guided by the rule. That is, the compatibilist holds the view that a particular instance of rule-guided choice is rational only if that choice is supported by the balance of reasons that apply to that case. end p.224 The third response is to adopt a revisionist approach. The revisionist agrees with the compatibilist that rule-guided behavior does have a real place within an instrumental theory of justification, but one that forces us to revise the standard account of practical reasoning. The revisionist, then, denies that if the rule gives the wrong result it would be irrational to follow it. A typical revisionist strategy is to argue for some sort of hierarchical model of justification, in which a distinction is made between justifying the choice of rules and justifying the choice of an action within the framework of the rules that have been adopted. 4 That means, however, that on the revisionist account, the door is opened to the possibility of rule-guided rational action that is not supported by the balance of reasons applicable at the time the action is taken. Predictably, whenever a revisionist insists that it is rational to be guided by a rule, the compatibilist will argue that the reasons for following the rule actually tip the balance of reasons in favor of conformity to the rule. Compatibilists are inclined, then, to argue that revisionism is not a coherent position, in that it must sanction actions that are not supported by the balance of reasons.
But revisionism has attracted a significant following in recent years. One can trace this, in great part, to the influence of Joseph Raz, who argued that it might be rational to act in ways that are not recommended by the balance of reasons. On Raz's view, rules can constitute “second-order reasons”—exclusionary reasons—not to act on certain firstorder reasons, where the latter are the sorts of reasons that lend themselves to the balancing approach. 5 When an exclusionary reason applies, the first-order reasons that are excluded by it are completely removed from the balance of reasons. Thus it would be possible that a balance of first-order reasons recommends an action but that the existence of an exclusionary reason requires the agent to disregard that recommendation and act in a different manner. Raz insists that certain reasons for acting must be exclusionary in nature, for if they were just first-order reasons, to be considered along with other reasons, they could not perform the function we normally attribute to them. Consider the example here that one should in certain contexts economize on deliberation costs by employing a rule of nondeliberation. 6 If this were just a first-order reason not to deliberate, then the only way of knowing if this reason were outweighed in a given context would be to consider all of the first-order reasons for action—that is, to deliberate—but that would defeat the very purpose of the rule of nondeliberation. Again rules that secure coordination can be understood as providing exclusionary reasons, for if everyone acted on the balance of reasons as they each saw it, concerted action would be impossible. Raz's work occasioned a lively exchange with a number of compatibilists. In part this constituted an in-house debate between Raz and a variety of other legal theorists. 7 But the dispute was more general. In part what the compatibilists argued was that any valid example of what Raz characterized as an exclusionary end p.225 reason could be treated simply as a first-order reason that was very weighty. It was also argued, however, that Raz had never really given an answer to the question of how one could be justified in not acting according to the balance of reasons. 8 Quite independently of the revisionist line of inquiry that Raz initiated, the issue of the rationality of rule-guided behavior arose also among economists and decision theorists. In the second half of the twentieth century, the prevailing view among economists was that the concept of a preference ordering, which was central to microeconomics, must be given a behavioral interpretation in terms of how the agent chooses. To say that one prefers x to y is to say that when confronted with a choice between x and y one will choose x. On this way of thinking, it is plausible to suppose that if one chooses to be guided by a rule, then in so doing one simply reveals one's preference for being so guided. To be sure, such a preference seems odd within the context of a theory of consumer choice. More to the point, appealing to the idea that in choosing to be guided by a rule one is simply revealing one's preference for being rule-guided has something of the air of a purely ad hoc assumption. However, the theory of decision making based on the maximization of one's preference ordering over possible options was taken over by decision and game theorists (see Joyce, chap. 8, Dreier, chap. 9, Bicchieri, chap. 10, and Danielson, chap. 22, this volume), who were happily prepared to expand the concept of a mathematically representable preference ordering (a utility function) to any decision
problem that an agent might confront. Within this context, then, it might seem as if the concept of rule-guided behavior posed no theoretically interesting problem. But in fact there were those who found themselves worried about this interpretive move. The most elegantly and powerfully formulated dissent came quite early on from the economist and philosopher Amartya K. Sen: A person is given one preference ordering, and as and when the need arises this is supposed to reflect his interests, represent his welfare, summarize his idea of what should be done, and describe his actual choices and behavior. Can one preference ordering do all these things? A person thus described may be “rational” in the limited sense of revealing no inconsistencies in his choice behavior, but if he has no use for these distinctions between quite different concepts, he must be a bit of a fool. (Sen 1977 , 336) Sen's disaffection with this sort of account is especially due to what he regarded as its lack of a coherent account of the notion of choice behavior based on commitment—just the sort of behavior that appears to be involved when one is led, in virtue of accepting a rule, to regard the rule as providing a reason for choosing one way rather than another. As it turns out, however, it was not just the facile assumption that commitment could be assimilated to a special kind of preference that engendered worries about how to understand rule-guided choice. In fact, the issue of the nature of rule-guided choice was forced upon economists by a special class of decision problems, in which one faced the far from anomalous problem that one's preferences now among various options that one would face in the future often turned out to be different from the preferences one would then have (that is, in the future) for those same options. That is, what brought the issue to a head was the problem of what came to be known as time-inconsistent preferences. The seminal work here is a 1955 article by R. H. Strotz, who argued that the characteristic way in which persons discounted the future lead to such time-inconsistent preferences. Strotz's own examples were primarily drawn from the area of personal decision making—as, for example, when one seeks to adopt a savings or dieting plan, only to find that, subsequently, one is tempted to depart from the plan. 9 But as Kydland and Prescott (1977 ) subsequently argued, the problem is also endemic to government policy making. 10
The striking thing about the ensuing analysis was that the standard theory of rational deliberation turned out to have only limited resources with which to deal with the issue raised. If one now wants that in the future one adheres to a rule, but one worries that the future self will want instead to disregard the rule, then on the standard way of thinking one has only one of two strategies available. One can precommit, that is, tie one's hands now so that one will be powerless to deviate from the rule in the future (a strategy that is a close cousin to that of agreeing to an enforcement mechanism that will compel one to behave a certain way in the future, upon pain of suffering an unacceptable outcome), or if one lacks the means to control one's future self in this manner, one can recognize now that planning to adhere to the rule in the future is futile, for it would be to adopt a plan that one realizes will never be executed. Such strategies, however, could be characterized as second-best strategies, for they require the self to forego adopting plans that would be beneficial, or to expend scarce resources to ensure that their future selves are bound to do their part. The problem of time-inconsistent preferences is closely related to another problem with which economists and game theorists have had to grapple, again with somewhat limited
resources, namely the free-rider problem. It is characteristic of many mutually beneficial arrangements between persons that what each participant stands to gain from is that others conform to the rules defining the arrangement, and that their own conformity to those rules is a cost to them. In such cases, each will be disposed to encourage others to conform but be disposed to avoid conforming whenever possible. The problem can arise, of course, even in the very simplest of contexts, such as contractual or reciprocal arrangements in which goods are exchanged, but there is a temporal separation between the performances expected of each. Both parties may be interested in the exchange, but the person who must perform first has to contend with the problem that if he or she performs, the other party may now have no motive to reciprocate. Once again, the standard theory of rationality has limited resources with which to deal end p.227 with this problem. As in the case of the self that has to deal with the rational dispositions of its own future self, those who have to deal with the rational dispositions of other selves are limited to either forgoing certain arrangements that would be mutually beneficial or expending scarce resources to ensure that others do their part—that is, by providing sanctions against free-riders. It is in this context that revisionist views could be said to have come into their own. One can start by noting first that Raz's exclusionary reasons account is applicable to cases of this sort. To commit to a personal rule binding one's own future self can be thought of as creating a second-order reason not to act on the first-order temptation one will subsequently come to have to depart from the rule. If such “temptations” were not excluded from consideration, the balance of first-order reasons—one's preference in the balance—might well be to depart from the rule; but given the exclusion of the “temptation” the balance of reasons may well favor adhering to the rule. Similarly, to the extent that a number of persons were able to approach cooperative situations from an exclusionary reasons perspective, they would be able to execute useful arrangements without having to employ expensive enforcement devices. Raz's exclusionary reasons account does not exhaust the resources of a revisionist view, however. There are, in fact, two other revisionist proposals that have more recently been made. One alternative involves appeal to what could be characterized as a constraint model, in which one thinks of commitment to a rule as a volitional act that alters the feasible set of future alternatives. Isaac Levi and Scott Shapiro have defended versions of this approach. 11 On this view, making a commitment to a rule involves an alteration in what one regards subsequently as the set of feasible alternatives from which choice is to be made. More specifically, making a commitment is to be understood as causally leading one to no longer regard certain actions as feasible. On this account, an agent is being (instrumentally) guided by a rule R if and only if (1) the agent conforms to R, (2) recognition that R applies constrains the agent to conform to R, making nonconformity infeasible, and (3) absent the constraint, the agent believes that he or she might not have conformed to R. The constraint model of rule-guided behavior has the great advantage that it offers an unproblematic account of the rationality of being guided by a rule, since rationality has always been taken as matter of how choice is to be made from among a feasible set of
alternatives. But the problem for this approach is how to explain what has to be regarded as a special sort of feasibility—how to account for the rational agent coming to regard, by an act of will, certain logically feasible (and attractive!) alternatives as no longer practically feasible. A third approach involves appeal to a resolute choice model. Both I and David Gauthier have explored versions of this approach. 12 On the resolute model, one is to suppose that—in contradistinction to the standard way of thinking, where present choice can only be adjusted to what one expects to choose in the future—a end p.228 rational agent can choose to constrain future choices to what one has decided in advance, to a prior commitment to a rule. On the version of this approach that I originally put forward, commitment is to be understood as a volitional act that (causally) transforms one's subsequent preferences, that is, one's future preferences regarding certain future actions. 13 This version of the resolute model, then, could be said to share with the compatibilist position the view that a particular instance of rule-guided choice is rational only if the balance of reasons applying in that particular case supports that choice. If the resolute model manages thereby to avoid the constraint model's problem of having to invoke a special notion of feasibility, it nonetheless faces a parallel problem: explaining how the commitment to a rule at one point in time can transform one's future preferences. It should be noted that there are a number of things that all three of these revisionist models have in common. First, they each view agents as capable of volitional acts that have temporal thickness, with the present self, at least in certain circ*mstances, being able to exercise rational authority over the behavior of the future self and thereby justifying future obedience to the rule adopted. Second, each of these ways of thinking about rule-guided choice preserves our sense that in such behavior there is a tension between commitments made and preferences at the moment of choice. On Raz's account, the temptation to break the rule, while excluded from deliberation, is still present. On the constraint model, there is still a tension between the commitment made and present preference, for the temptation to break the rule has once again not been removed; it is just that it is no longer viewed as a feasible option. On the resolute choice model, even though one's overall preference, at the subsequent point of choice, is for conforming to the rule, this does not mean that one is not tempted to break the rule—for were other things equal (which they are not, since one has made a commitment) one would prefer to break the rule, and the awareness of this fact preserves the sense of a tension between commitment and choice. Third, in each case, as the above examples make clear, the analysis need not be confined to cases of personal rules. There is, on each of the three accounts, a natural extension from the case of the agent faced with making decisions over time to the case of different agents trying to coordinate their actions. 14 From an analytic point of view, one important issue is obviously whether one or another of these revisionist models can be shown to be superior to the others, in respect to giving a satisfactory account of rule-guided choice. And there is also an unresolved question whether these three models exhaust the possibilities. The work of Ainslie on personal rules seems to combine elements of both the constraint and the resolute models, but this may not do justice to what is a very original and interesting account of personal rules. 15
Ainslie's work also brings to the subject extensive empirical findings drawn from the fields of physiology and psychology, and it is clear that future analysis could greatly benefit from more integration of such empirical work with the philosophical, legal, and decisionend p.229 theoretic perspectives that I have been discussing. So also there is much to be gained by examining the recent work of institutional theorists, work that has greatly enriched our understanding of the role of rules in social, political, and economic activities. 16 I want to put the issue of the best way to characterize rule-guided choice to one side, however, for I think there is an even more pressing problem. This is a problem that arises in virtue of an objection that can be raised to any of the models discussed by adapting an argument of Michael Bratman's regarding the role that plans or intentions play in our deliberations. Intuitively, we want to say that intending to do something provides a reason for action. But one has to be careful here. Bratman (1987 , sec. 3.3) argues that, as stated, the view is too strong. Even though the idea of an intention-based reason allows us to take seriously the way in which one sees one's prior intentions as directly relevant to the conclusions of one's further practical reasoning, it appears to sanction an unacceptable form of bootstrapping. Suppose that at some subsequent point in time, one comes to reconsider one's previous intentions. It would be an unacceptable form of bootstrapping, Bratman argues, to suppose that the mere fact that one previously intended (as distinct from one's evaluation now of the consequences of acting on that intention) carried any weight. Adapting this point to the present discussion of a commitment to rules, the issue is how to understand the role such a commitment plays in deliberation without falling into the unacceptable position of letting the mere fact of having made the commitment create a distinct reason in favor of guiding one's choices by reference to the rule. Bratman thinks that the way to avoid such bootstrapping is to treat the adoption of a plan not as creating new preferences, but as merely constraining the range of actions available for subsequent choice. And Gauthier (1996 , 218), in commenting upon Bratman, thinks that this is the way to deal with the bootstrapping problem. This would seem to suggest that each is implicitly inclined to think that a constraint model is the most appropriate revisionist model. It can be noted, moreover, that in Gauthier's own explorations of “resolute choice” he clearly dissents from the suggestion, in my book, that the best way to understand resolute choice is by appeal to the idea of a commitment that causes one to have different preferences for the options available at some subsequent choice point—to the idea of what I characterized there as “endogenous preference changes.” 17 My own sense, however, is that the issue goes deeper. It is not just a problem that arises within a resolute model of commitment, that is, one that invokes the idea of endogenous changes in preference. The issue is really one of intelligible reasons, and Raz's exclusionary reasons model, as well as Levi and Shapiro's constraint model, strike me as posing the same sort of problem. That is, the more general question is just how the agent's commitment to a rule could provide reasons for the agent to be guided by that rule. 18 The problem is that it seems odd to suppose, within an instrumental framework, not only that a commitment end p.230
to a rule could create a new situation in which the balance of preferences would be tipped in favor of adhering to the rule, but also that it could yield a revised view of what is feasible or generate some hierarchical structure of first- and second-order reasons, with the latter constituting exclusionary considerations. Bratman's bootstrapping worry is, I think, at root a worry about how to provide a role for rules within an instrumental theory of rationality without finding oneself faced with a “bootstrapping” move that leaves one advocating “rule worship” or “rule fetishism.” 19 Why should we suppose that telling the story in terms of constraints on the feasible set, or in terms of second-order reasons, rather than in terms of (new) preferences over those same options, somehow solves the problem of bootstrapping? The key to understanding how to avoid the bootstrapping charge lies, I suggest, not in how we characterize commitment to a rule, but in what story we can tell about what drives or motivates the commitment. And here it will prove useful to return and explore one paradigm case of rule-guided choice—namely situations in which the utilization of a rule of nondeliberation (i.e., a rule of nondeliberative nonreconsideration) serves to reduce decision-making costs. 20 I noted above that Raz's exclusionary reasons approach can be illustrated by this kind of case. The argument was that if this were just a first-order reason not to deliberate, then the only way of knowing if this reason were outweighed in a given context would be to consider all of the first-order reasons for action, that is, to deliberate—but that would defeat the very purpose of the rule of nondeliberation. The argument, then, is that there is a cost-savings consideration that drives the commitment, savings that cannot be achieved except by not being disposed to reconsider continually the pros and cons of applying the rule in each subsequent case that arises. Having such a rule of nonreconsideration in place for certain classes of cases will, then, clearly save on costs of deliberating. Of course, as Bratman makes clear, any such rule will be subject to all sorts of exceptions. That is, the savings in deliberating costs can be outweighed by the severity of the negative consequences that can flow from not deliberatively reconsidering in certain contexts. Presumably, then, the commitment to nondeliberative nonreconsideration will not be unconditional: certain conditions will serve to trigger deliberative reconsideration. But invoking Raz's notion of economizing on decision-making costs by having a second order exclusionary rule is not the only way one can rationalize a rule of nondeliberate nonreconsideration. We could also appeal to a constraint model and characterize being guided by such a rule as a case in which the option of reconsideration, which is typically available to us at subsequent points in time, is simply (for at least certain classes of cases) treated as a nonfeasible option. That is, we could interpret commitment to such a rule in terms of the constraint model. Similarly, one could interpret commitment to such a rule as a case of acting in a resolute fashion—that is, deciding, for reasons again of cost savings, to adopt a rule of nonreconsideration, which would lead one to prefer to not deliberatively reconsider some decision reached previously, unless, say, the real possibility of significant negative consequences was thrust upon one. Whichever story is told, it is clear that each of these ways of describing the rationality of a rule of nondeliberative nonreconsideration commits one to a revisionist program. That is, to accept such a rule and be guided by it is to put oneself in a position in which the
balance of reasons could turn out, if only one had deliberatively explored them, to favor reconsideration. That is, the rule counsels one to choose in a manner that will not always ensure that one chooses in accordance with the balance of reasons that arise within the context of a particular act of choice. Thus accepting such a rule cannot be rationalized within the framework of a compatibilist position. Moreover, whichever one of the three ways of interpreting a revisionist account one employs, in each case it would appear possible to avoid the bootstrapping charge. This is because the commitment to the rule, on any of these accounts, is based on a pragmatic consideration—on a consequentialist concern to hold down costs. 21 What drives the argument, then, is not the mere fact of making a commitment to the rule of nonreconsideration, but the cost-saving consideration behind the making of that commitment. Now, against the background of these considerations, consider once again the way in which the standard theory of rationality deals with the self over time (and with persons who have to interact sequentially with one another). 22 First, the standard theory is really distinguished by a particular view of what it means for the self to deal with its own future selves, and, correspondingly, for the person to deal with other persons. This view, I want to argue, is essentially an autistic view of how the choice behavior of other selves or other persons is to be worked into the analysis. On this way of thinking, the choices made by one's own future self, or other persons, constitute so many conditioning factors— variables—whose value a rational agent must try to estimate, not for the sake of coordinating one's own choices with those other choices, but simply with a view to enabling the present self vis-à-vis future states of the self, and the self vis-à-vis other persons, to maximize its own utility function. That is, the perspective that is presupposed is that these conditioning factors are to be treated in the same way that other factors— states of nature—are to be treated: as variables whose value must be properly estimated if one is to achieve effectively maximum preference satisfaction. These are models, then, in which the self ends up taking the choice behavior of the future self (or other selves) as given, as something to which one's present self must adjust. In the case of sequential choice, however, the standard way of conceiving the constraints on a prior decision maker are a bit more complicated. This is because the later self (or the other person who chooses subsequently) can be expected to maximize from its own perspective, against the known choice of the earlier self or person. The problem for the earlier self, then, is that it must contend with a end p.232 subsequent self that will not have to make a choice until the earlier self has chosen, and will then be in a position to maximize from its own perspective. Thus, the person who, in the morning, resolves to adhere strictly to a diet has to face the prospect that he or she will, in the evening, have a preference for taking a helping of some forbidden dessert. Correspondingly, a parallel problem is posed for two farmers, A and B, where A's crops will be ready for harvest earlier than B's. 23 B would be willing to help A first, if only, subsequently, A would reciprocate. But B has to contend with the real possibility that A will, when the time comes, refuse to reciprocate. Given this, the task of the earlier self (or first person to offer to help) is to choose so that the final outcome, given that the
subsequent self (or other person) maximizes against what the earlier self (or person) has done, is maximally preferred by the earlier self (or the first to offer to help). Thus, the standard model locks us into a view of mutual maximizing adjustments of each to the other, with the additional stipulation that given the existence of a natural temporal order, the process of mutual adjustment is structured in a certain way: the later self (or person) merely reacts to the earlier self (or person); the earlier self (or person) reacts to the predicted reaction of the subsequent self (or person). 24 Now it is a well-established fact that such a form of strategic interaction between earlier and later selves, or between two persons who choose sequentially, will typically generate an outcome that is less advantageous for each of the selves involved than one that would be possible to achieve if the selves involved were to coordinate their choices cooperatively. In the language of the economist, the outcome will be Pareto-suboptimal. 25 To return to our examples, the would-be dieter must employ some device to “tie the hands” of his subsequent self—a device whose employment means that his goal is achieved in a more costly manner (than if he could simply resolve to keep to his diet). Correspondingly, the two farmers will either fail to help each other, to the detriment of both, or they will have to expend additional resources to ensure in some fashion or other conditions of trust. The farmer's problem, then, is simply a sequential version of the standard Prisoners' Dilemma game, in which one player chooses and then the other player chooses in full knowledge of how the first has chosen. 26 One can escape from the resultant suboptimality of the outcome only by replacing the kind of strategic interaction just described with something radically different, by each self approaching the sequential decision problem as a coordination problem. 27 That is, effective coordination— coordination that promotes the interests of both—will require that each restrain her choices and not straightforwardly maximize. It would be misleading, then, to suggest that the revisionist models turn upon the idea that commitment to a rule is just a regimentation of subsequent choice to an earlier choice of a rule. What we really seek is a model in which regimentation in either direction is replaced with some sort of cooperative form of interaction. More specifically, what the resolute model (or any other viable revisionist end p.233 model) must make clear is that rational interaction can and should proceed on other grounds than each maximizing against what the other self or person has chosen or can be expected to choose. That is, what could rationalize an alternative form of interaction would be precisely the possibility of the mutual gain to be achieved thereby. But in this very consideration of mutual advantage, I suggest, one can find just the kind of practical perspective that could motivate the parties—the sequence of selves or the different persons—to make a commitment to act in a rule-guided manner. 28 If we suppose that rational selves or persons will be disposed to coordinate when so doing works to their mutual advantage, we have an account of how such selves or persons could make a commitment to a rule. And this will be a commitment that is fully consistent with a practical or instrumental perspective. The later self (or the person who chooses subsequently) can be understood to choose in a rule-guided fashion because this is a strategy from which it, no less than the prior self (the person who chooses first), stands to
benefit—even though, of course, as the argument goes from the perspective of the standard theory, the subsequent self (or person) has a reason to depart from the rule at that point. The point is that even though the subsequent self (or person) could do even better by departing from the rule, the level of gain to be achieved by nonetheless adhering to the rule is greater than what the subsequent self (or person) could have hoped to have achieved in the absence of its willingness to be guided by the rule. 29 Its reason for being guided by a rule in the choice it makes is, then, a thoroughly practical, instrumental reason: absent such rules, which serve to coordinate choice, each party to the interaction stands to do less well. 30 What is invoked in such a way of thinking is not some bootstrapping move in which commitment to a rule makes us liable to objectionable rulefetishism, but rather simply a more global or holistic way of thinking about consequential benefits. 31 To see this, compare the following two possible worlds. In the first, each of the members of some society of persons is disposed to cooperate with one another as a matter of principle—that is, they do so without having to employ any system of mutual surveillance and enforcement. For the sake of the argument, we may simply suppose that in the society in question there is a cultural tradition to this effect. Now imagine a second possible world just like the first, but where such cooperation is possible only so far as resources are expended on surveillance and enforcement devices that serve to provide persons with sufficient incentive not to “free ride” on the cooperative efforts of others, and thus to stabilize the cooperative arrangement. On the assumption that the costs of such devices are shared among all, it is plausible to suppose that everyone is better off in the former than in the latter world. In each, cooperation is ensured, but in the latter only by the expenditure of resources that do not have to be expended in the former. To be sure, in the former world one might imagine that some come to realize that if they put to one side their cultural habits, they could do even better. But that end p.234 world is not stable. It will be rational for those who are thus taken advantage of to agree to expend resources to bring the free-riders into line. Of course, so long as most continue to cooperate as a matter of principle, the surveillance and enforcement costs will be less than they would be in a world in which all require such inducements. But so long as the inducement system is reasonably effective, and costs are shared, all would still do better to participate in the first possible world described above than in this modified version of that world. What remains, now, is simply to ask, given the consequential superiority of the first possible world to the alternatives in question, whether rational beings might not deliberatively adopt such an arrangement. On the standard theory the answer is “No!” But does it really make sense to suppose that the benefits that each stands to realize by living in the first possible world can be secured only when, as a matter of pure cultural history, such cooperative norms are already in place—that it cannot be something that rational beings could bring about by any sort of deliberative process? That the standard theory must deny this possibility seems to me a telling argument against it—and one that proceeds by citing what are clearly consequential considerations. 31
As this last remark is designed to suggest, if we now return once again to the issue of bootstrapping, it should be clear that this sort of way of understanding rule-guided choice is not subject to Bratman's criticism. However we choose to interpret being guided by a rule—whichever of the three revisionist models we employ—the manner in which being rule-guided is rationalized makes it clear that there is no bootstrapping going on. It is interesting to note, moreover, that the case we previously discussed, in which one is guided by a rule of nondeliberative nonreconsideration, can be fully rationalized in terms of this expanded and more holistic way of thinking about consequences. Indeed, were the costs to be saved by the adoption of such a rule credited only to the previous self, one could complain that this provides an inadequate pragmatic rationale for the adoption of such a rule. But the interesting thing about cost savings that the self can realize over time is that they are savings that can be carried forward and thus benefit the later self as well as the earlier self. Similarly, the cost savings that can be achieved by persons being able to coordinate their choices without having to employ expensive surveillance and enforcement devices are savings that not only can typically be shared among the participants, but plausibly must be shared, if what is clearly a much more rational form of interaction is to be realized.
NOTES In preparing this article I have adapted a certain amount of terminology and framing of points from an encyclopedia entry that Scott Shapiro and I jointly authored (Mc end p.235 Clennen and Shapiro 1998 ). When utilizing material from a joint undertaking, one is bound to be especially concerned to acknowledge when the borrowed material is primarily the work of the other author. Let me just say, in respect to the material I have borrowed, that the elegant precision with which many of the points and phrasings were made convinces me that they must have been due to Scott, and I am grateful, then, for his having given me permission to draw from what is officially our joint work. 1. The argument to be explored here could easily be adapted to address the larger issue of the place of commitment in a theory of rationality, as posed, for example, in a seminal article by Amartya K. Sen (1977 ). I shall content myself with developing the argument as it applies to the more special case of making a commitment to be guided by a rule. 2. A version of this argument is sometimes invoked to show that rule-utilitarianism must collapse into act-utilitarianism. Conversely, the line of argument I shall be exploring in this paper implies that that rule-utilitarianism is, at least with respect to this consideration, a perfectly coherent theory. I shall not explore this implication here, however, primarily because I think there are other substantial objections that can be raised against rule-utilitarianism. 3. Goodwin was perhaps the first to embrace this conclusion, in Goodwin (1793 ) 1971 . More recently, Wolff opted for a similar position in Wolff 1968 . 4. See Rawls 1955 , Wasserstrom 1961 , and Hooker 2000 .
5. This theme finds expression in a number of Raz's works, including Raz 1979 , 1986 , and 1990 . 6. Cases of this sort are also analyzed with much care and insight in Bratman 1987 . 7. See, for example, Green 1988 , Moore 1989 , Alexander 1990 , and Hurd 1991 . Raz offers a response to some of his critics in Raz 1989 . 8. See Moore 1989 and Schauer 1991 . 9. This kind of case was also the focus of a series of articles by Thomas Schelling. See Schelling 1984a , 1984b , and 1985 . It is also the focus of Elster 1984 and Ainslie 1988 , 1992 , and 2001 . 10. See also Yaari 1977 for a very comprehensive analysis of the logical structure of such problems. 11. See in particular, Levi 1986 . Scott Shapiro, who studied with Levi, developed this idea at great length in his dissertation. See Shapiro 1996 ; see also McClennen and Shapiro 1998 . 12. The term “resolute” was first used in this way in McClennen 1985 . I subsequently returned to this topic in a number of other articles, including McClennen 1988a and 1990 . McClennen 1988b is devoted to exploring the relation between this idea and Gauthier's idea of constrained maximization, as developed in Gauthier 1986 . In McClennen 1993 I take up the question of the relation between the idea of being resolute and the conception of planning developed in Bratman 1987 and 1992 . In McClennen 1997 I try to develop a less formal conception of resolute choice than the one contained in my book. During the late 1980s and 1990s, however, Gauthier himself contributed significantly to the analysis of being resolute in a series of articles. See, in particular, Gauthier 1988–89 , 1990 , 1994 , 1996 , 1997a , 1997b , 1998a , and 1998b . What is especially important about Gauthier's own explorations is that he has, I think, made it clear that the resolute model may not apply to games involving threats rather than assurances, and that even in the case of assurance games, there are limits to the model that must be acknowledged when chance events are also involved. My sense is that these limitations may also have to be acknowledged in the case of the other two revisionist models that I have been discussing. 13. See McClennen 1990 , sec. 12.7. 14. Indeed, the possibility of moving back and forth between the case of adopting a personal rule that is to be regulative of ones future choices and cases in which the rule in question serves to coordinate the choices of distinct individuals is implicit in the economist's very conceptualization of deliberation. To suppose that the earlier and the later self have different preferences—and hence different utility functions—is to suppose that the deliberative problem is essentially that between two different individuals who must interact in some fashion or other with one another. 15. For a brief introduction to Ainslie's account, see Ainslie 1988 . Ainslie 1992 and 2001 contain much more detailed accounts. 16. See here, in particular, Brennan and Buchanan 1987 and North 1990 . 17. McClennen 1990 , 214–15. 18. Recall that the issue here is how to understand such revisionist views within the context of an instrumental account of reasoning. That is, the issue is not whether it is possible to enjoy simply acting in conformity with a rule, or to attach intrinsic value to so acting. 19. This issue is explicitly raised in Bratman 1992 .
20. The terminology here is from Bratman 1987 , chap. 5. 21. I am using “consequentialist” in its most catholic sense here (not in the very narrow sense in which many twentieth moral philosophers have used it), to describe the position that all the consequences that flow from an action, for all the persons involved, are to be brought, and brought equally, into consideration—thereby invoking some version or other of a utilitarian rule of the maximization of the sum of benefits. On the other hand, I do suppose that if, on other grounds, some version of classical utilitarianism were defensible, the argument of this paper could be brought to bear in support of the idea of a rule version of that utilitarianism. 22. In what follows I will concentrate on the case in which cooperation with other persons involves a series of sequential acts, rather than the case where all choose simultaneously. But the argument to be developed can be extended to the case of simultaneous choice as well. 23. See Gauthier 1994 , 692. 24. In the literature on game theory this gives rise to the requirement that the sequence of choices made by rational persons (or rational time-defined moments of an individual person) satisfy what is known as the subgame-perfect equilibrium condition. For an elementary discussion, see Kreps 1990 , chap. 4. In McClennen 1990 , sec. 15.1, I argue that this condition is forced upon us by a separability assumption that mandates viewing the decision problem at any point in a decision tree as if it were a new decision problem involving only what is still open to agents to choose from that point onward. It is just this condition that gives rise to the idea that sequential choice problems can be analyzed by backward induction—by solving the last stages of a sequential problem first and then moving progressively backward through the tree until one comes to the first point at which choice is to be made. My own account of resolute choice was designed to break what I regarded as the stranglehold that this separability assumption and the related notion of backward induction had on the analysis of sequential choice problems— end p.237 a stranglehold that generates the profoundly implausible conclusion, for example, that under conditions of common knowledge, when two persons play a Prisoners' Dilemma game a known finite number of times, the only rational solution is for each to confess each and every time! The role that the separability assumption plays is discussed in length in McClennen 1997 . An especially lucid and accessible discussion of this assumption is also to be found in Gauthier 1996 . 25. Suboptimality in the case of the self over time means that each and every timedefined self is less well off than it would have been under some alternative arrangement. Both Kydland and Prescott 1977 and Yaari 1977 emphasize the suboptimality of the arrangements that can be realized within the framework of the standard theory of rationality. 26. If one considers the problem of two or more persons choosing simultaneously rather than sequentially, a similar result obtains. In terms of the standard theory of rational deliberation, the subgame perfect equilibrium condition is now replaced by the Nash equilibrium condition, and players who are fully rational (in the standard sense) will end up at an equilibrium outcome that is suboptimal, as the traditional version of Prisoners'
Dilemma makes clear, whether they simply do not trust each other and choose accordingly, or they expend resources to establish conditions of trust. 27. On the other hand, only in very special cases will such “games” constitute pure coordination games in which there is no conflict of interest. When the “game” is of this type, of course, it will have a trivial solution in virtue of its sequential structure. What makes for a pure coordination problem is that there are multiple solutions that are equally good, and the parties are unsure of which one to coordinate upon. But given a sequential structure, the party that goes first can thereby provide an appropriate signal as to which point to coordinate upon. 28. The distinction here would seem to mirror the distinction, to be found in Atiyah 1989 , between taking a contract as binding simply in virtue of one's having given one's word (simply made a commitment) and in virtue of the benefit that one expects to derive thereby. In opting for the latter conception, I am supposing that one's commitment itself derives from some consideration and does not just arise as a result of a pure act of will. And that also means that the consideration is relevant to the issue of whether the subsequent self is bound to the arrangement. 29. Because one still foregoes the additional gain that could be realized by deviating from the rule, the balance of reasons could still be said to support such a deviation; to appeal, then, to this rationale for being guided by rules is to appeal to a revisionist account. 30. This, one can insist, is a very different consequentialist consideration than, say, the consideration that one risks punishment or other losses if one is discovered to have not adhered to the rule. Again, the reader needs to bear in mind that I am using “consequential” in a broad sense: the argument does not turn on invoking the utilitarian ideal that the sum of benefits to all those who participate is greater. 31. Elster (1984 ) briefly alludes to just such a more holistic or global way of thinking about benefits, as it would apply to the problem of the self over time. The self that does not always, at each point in time, take the direction that leads upward will be able to avoid the problem of local gradient maxima in his or her path that are inferior to the gradient maxima to be found further on. That is a consequentialist consideration or reason in support of being, in effect, resolute. However, Elster retreats from the ex end p.238 pected conclusion that rational agents should adopt such an approach, arguing instead that such self-restraint is not something that (imperfect?) human deliberators can achieve. But he fails to explain just why this lies beyond the power of mortals. Thus he argues that Ulysses is less than perfectly rational, because he has to resort to the device of having himself tied to the mast. Still, he goes on to insist, “he is capable of achieving by indirect means the same end as a rational person could have realized in a direct manner” (Elster 1984 , 36). The implication remains, it would seem, that he is not fully rational because he can achieve his end (avoiding the Sirens and continuing on home) only in a more costly manner. But, then, one must insist: What holds Ulysses back from the more efficient approach? end p.239
chapter 13 MOTIVATED IRRATIONALITY Alfred R. Mele The literature on motivated irrationality has two primary foci: action and belief. In the former sphere, akratic action—action exhibiting so-called weakness of will or deficient self-control—has received pride of place. Philosophical work on motivated irrational belief includes, but is not limited to, work on self-deception. The primary topics of this chapter are akratic action and motivationally biased belief.
1. Akrasia, Self-Control, and Strict Akratic Action The classical Greek term akrasiais formed from the alpha privative and kratos—strength or power. The pertinent power is the power to control oneself. Hence, akrasiais deficient self-control. Self-control, in this sense, is, very roughly, a robust capacity to see to it that one acts as one judges best in the face of actual or anticipated competing motivation. 1 The trait may be either regional or global (Rorty 1980a ). A scholar who exhibits remarkable self-control in adhering to the end p.240 demanding work schedule that she deems best for herself may be akratic about smoking. She is self-controlled in one region of her life and akratic in another. Self-control also comes in degrees: some self-controlled individuals are more self-controlled than others. People with global self-control—self-control in all regions of their lives—would be particularly remarkable, if, in every region, their self-control considerably exceeded that of most people. In Plato's Protagoras,Socrates says that the common view about akratic action is that “many people who know what it is best to do are not willing to do it, though it is in their power, but do something else” (Plato 1953 , 352d). Here he raises (among other issues) the central question in subsequent philosophical discussion of akrasia:Is it possible to perform uncompelled intentional actions that, as one recognizes, are contrary to what one judges best, the judgment being made from the perspective of one's values, principles, desires, and beliefs? More briefly, is strictakratic action possible (Mele 1987 , 7)? Relevant judgments include judgments in which “best” is relativized to options envisioned by the agent at the time. In strict akratic action, an agent need not judge that a course of action Ais the best of all possible courses of action open to her then. She may judge that Ais better than the alternatives she has envisioned—that it is the best of the envisioned options. If, nevertheless, in the absence of compulsion, she does not Aand intentionally pursues one of the envisioned alternatives, she acts akratically. It is a truism that a perfectly self-controlled agent would never act akratically. So akratic action, if it is possible, exhibits at least imperfect self-control. 2 The judgment against which an agent acts in strict akratic action is what I call a decisivejudgment. An agent's judgment that Ais the best of his envisioned options at a
given time is a decisive judgment, in my sense, if and only if it settlesin the agent's mind the question of which member of the set is best (from the perspective of his own desires, beliefs, etc.)—and best not just in some respect or other (e.g., financially), but without qualification. 3 Ann judges that A-ing would be morally (or aesthetically, or economically) better than B-ing and yet, in the absence of compulsion, she intentionally B-s rather than A-s. In B-ing, Ann need not be acting akratically, for she may also judge, for example, that all things considered,B-ing would be better than A-ing. A feature of paradigmatic strict akratic actions that typically is taken for granted and rarely made explicit is that the judgments with which they conflict are rationallyformed. In virtue of their clashing with the agent's rationally formed decisive judgment, such actions are subjectively irrational (to some degree, if not without qualification). There is a failure of coherence in the agent of a kind directly relevant to assessments of the agent's rationality. 4 The occurrence of strict akratic actions seems to be an unfortunate fact of life. Unlike many such (apparent) facts, however, this one has attracted considerable philosophical attention for nearly two and a half millennia. A major source of the interest is obvious: strict akratic action raises difficult questions about connection between evaluative judgment and action, a connection of paramount importance for any theory of the explanation of intentional behavior that accords evaluative judgments an explanatory role. Matters are complicated by our having—both in various theoretical approaches to understanding action and in ordinary thought—a pair of perspectives on the explanation of intentional action, a motivationaland an intellectualone (Mele 1995 , 16–19; Pettit and Smith 1993 ). Central to the motivational perspective is the idea that what agents do when they act intentionally depends on where their strongest motivation lies then. 5 This perspective is taken on allintentional action, independently of the species to which the agents belong. If cats, dogs, and human beings act intentionally, the motivational perspective has all three species in its sights. The intellectual perspective applies only to intellectual beings. Identifying minimally sufficient conditions for membership in the class of intellectual beings is an ambitious task, but it is clear that the work of practical intellect, as it is normally conceived, includes weighing options and making judgments about what it is best, better, or “good enough” to do. Central to the intellectual perspective is the idea that such judgments play a significant role in explaining some intentional actions. Many philosophers seek to combine these two perspectives into one in the domain of intentional human action. One tack is to insist that, in intellectual beings, motivational strength and evaluative judgment always are mutually aligned. Socrates seemingly advances this view in connection with his thesis that people never knowingly do wrong (Plato 1953 , 352b–358d). Theorists who take this tack have various options. For example, they can hold that judgment causally determines motivational strength, that motivational strength causally determines judgment, or that judgment and motivational strength have a common cause. They can also try to get by without causation, seeking purely conceptual grounds for the alignment thesis. The apparent occurrence of strict akratic actions is a problem for this general tack. The motivational perspective is well suited to akratic action: when acting akratically, one presumably does what one is most strongly motivated to do then. But the intellectual
perspective is threatened; more precisely, certain interpretations of, or theses about, that perspective are challenged. In threatening the intellectual perspective while leaving the motivational perspective unchallenged, akratic action poses apparent difficulties for the project of combining the two perspectives into a unified outlook on the explanation of intentional human action. That is a primary source of perennial philosophical interest in akratic action. There is much to recommend the motivational and intellectual perspectives, and a plausible combination is theoretically desirable. To some theorists, the threat that strict akratic action poses to a unified, motivational/intellectual perspective seems so severe that they deem such action conceptually or psychologically impossible (Hare 1963 , chap. 5; Pugmire 1982 ; Watson 1977 ). 6 Many others try to end p.242 accommodate strict akratic action in a unified perspective (Davidson 1980 , chap. 2, 1982 ; Dunn 1987 ; Mele 1987 , 1995 ; Pears 1984 ).
2. Explaining Strict Akratic Action To the extent that one's decisive judgment derives from one's motivational attitudes, it has a motivational dimension. 7 That helps explain why many regard akratic action as theoretically perplexing. How, they wonder, can the motivation associated with a judgment of this kind be outweighed by competing motivation, especially when the competing motivational attitudes—or desires,broadly construed—have been taken into account in arriving at the judgment? Elsewhere (Mele 1987 ) I defended an answer to this question that rests partly on two theses, both of which I defended. 1.Decisive judgments normally are formed at least partly on the basis of our evaluation of the “objects” of our desires (i.e., the desired items). 2.The motivational force of our desires does not always match our evaluation of the objects of our desires. (Santas 1966 , Smith 1992 , Stocker 1979 , Watson 1977 ) 8 If both theses are true, it should be unsurprising that sometimes, although we decisively judge it better to Athan to B,we are more strongly motivated to Bthan to A.Given how our motivation stacks up, it should also be unsurprising that we Brather than A. Thesis 1 is a major plank in a standard conception of practical reasoning. In general, when we reason about what to do, we inquire about what it would be best, or better, or “good enough” to do, not about what we are most strongly motivated to do. When we ask such questions while having conflicting desires, our answers typically rest significantly on our assessments of the objects of our desires—which may be out of line with the motivational force of those desires, if thesis 2 is true.
Thesis 2, as I argued in Mele 1987 , is confirmed by common experience and thought experiments and has a foundation in empirical studies. Desire-strength is influenced not only by our evaluation of the objects of desires, but also by such factors as the perceived proximity of prospects for desire-satisfaction, the salience of desired objects in perception or in imagination, and the way we attend to desired objects (Ainslie 1992 , Metcalfe and Mischel 1999 , Rorty 1980a ). Factors such as these need not have a matching effect on assessment of desired objects. A few hours ago, an agent decisively judged it better to Athan to B,but he end p.243 now has a stronger desire to Bthan to A.Two versions of the case merit attention. In one, along with the change in desire strength, there is a change of judgment. For example, last night, after much soul-searching, Al formed a decisive judgment favoring not eating after-dinner snacks for the rest of the month and desired more strongly to forego them than to indulge himself; but now, a few hours after dinner, Al's desire for a snack is stronger than his desire for the rewards associated with not snacking, and he decisively judges it better to have a snack than to refrain. In another version of the case, the change in relative desire strength is not accompanied by a change of judgment. Al retains the decisive judgment favoring not eating after dinner, but he eats anyway. Assuming that Al eats intentionally and is not compelled to eat, this is a strict akratic action. Empirical studies of the role of representations of desired objects in impulsive behavior and delay of gratification (reviewed in Mele 1987 , 88–93; see Mischel et al. 1989 for an overview) provide ample evidence that our representations of desired objects have two important dimensions, a motivational and an informational one. Our decisive judgments may be more sensitive to the informational dimension of our representations than to the motivational dimension, with the result that such judgments sometimes recommend courses of action that are out of line with what we are most strongly motivated to do at the time. If so, strict akratic action is a real possibility—provided that at least some intentional actions that conflict with agents' decisive judgments at the time of action are not compelled. A discussion of compulsion would lead quickly to the issue of free will, which is well beyond the scope of this chapter. It is worth noting, however, that unless a desire is irresistible, it is up to the agent, in some sense, whether she acts on it. This idea is an element of both the motivational and the intellectual perspective on intentional action. Another element is the idea that relatively few desires are irresistible. Of course, a proper appreciation of the latter idea would require an analysis of irresistible desire. 9 It may suffice for present purposes to suggest that, often, when we act against our decisive judgments, we could have used our resources for self-control in effectively resisting temptation. 10 Normal agents can influence the strength of their desires in a wide variety of ways (Ainslie 1992 , Metcalfe and Mischel 1999 , Mischel et al. 1989 ). For example, they can refuse to focus their attention on the attractive aspects of a tempting course of action and concentrate instead on what is to be accomplished by acting as they judge best. They can attempt to augment their motivation for performing the action judged best by promising themselves rewards for doing so. They can picture a desired item as something unattractive—for example, a chocolate pie as a plate of chocolate-coated chewing
tobacco—or as something that simply is not arousing. Desires typically do not have immutable strengths, and the plasticity of motivational strength is presupposed by standard conceptions of self-control. Occasionally we do notact as we judge best, but it is implausible that, in all such cases, we cannot end p.244 act in accordance with these judgments. (This suggestion is defended in Mele 1987 , chap. 2; 1995 , chap. 3; and 2002 .) 11
3. Kinds of Akratic Action and the Irrationality of Strict Akratic Action Not all akratic action is of the strict kind. In this section, without aspiring to be exhaustive, I identify two additional kinds discussed in the literature. I also comment on the question of whether strict akratic action is necessarily irrational. Socrates, in defending the thesis that no one knowinglydoes wrong, argues that what actually happens in apparent instances of strict akratic action is that, owing to the proximity of anticipated pleasures, agents change their minds about what it would be best to do (Plato 1953 355d–357d). Even if he mistakenly denies the reality of strict akratic action, Socrates identifies an important phenomenon. Some such changes of mind are motivationally biasedprocesses (on motivated bias, see secs. 4–6). When one is tempted to do something that conflicts with one's decisive judgment, one has motivation to believe that the tempting option would be best. After all, acquiring that belief would diminish one's resistance to acting as one is tempted to act (Pears 1984 , 12–13). In what I elsewhere (Mele 1996 ) called “Socratic akratic action” (without meaning to suggest that Socrates regarded the episodes as akratic), the agent's new belief issues from a process biased by a desire for the tempting option. In the context of akrasia,the most relevant standards for determining bias are the agent's. Perhaps the agent accepts a principle about beliefs that is violated by his present change of mind—for example, the principle that it is best not to allow what one wants to be the case to shape what one believes is the case. And if a motivationally biased change of mind of this kind is avoidable by the agent by means of an exercise of self-control, it is itself an akraticepisode, an episode manifesting akrasiaor an associated imperfection, for no perfectly self-controlled person makes motivated judgments that are biased relative to his own standards, if he can avoid doing so by exercising self-control. Furthermore, an intentional action that accords with the new judgment is derivatively akratic (Mele 1987 , 6–7, 1996 ; Pears 1984 , 12–13; Rorty 1980b ). Seemingly, there also are “unorthodox” instances of akratic action, in which agents act in accordance withtheir decisive judgments, and, similarly, unorthodox exercises of selfcontrol in support of conduct that conflicts with the agents' decisive judgments (Bigelow, Dodds, and Pargetter 1990 , 46; Hill 1986 , 112; Jackson end p.245
1984 , 14; Kennett 2000 , 120–24; Mele 1987 , 7–8, 1995 , 60–76). Here is an illustration of unorthodox akratic action from Mele 1995 : “Young Bruce has decided to join some wayward Cub Scouts in breaking into a neighbor's house, even though he decisively judges it best not to do so.At the last minute, Bruce refuses to enter the house and leaves the scene of the crime. His doing so because his decisive judgment has prevailed is one thing; his refusing to break in owing simply to a failure of nerve is another. In the latter event, Bruce arguably has exhibited weakness of will: he `chickened out'” (60). If, instead, Bruce had mastered his fear and participated in the crime, we would have an unorthodox exercise of self-control. John Bigelow, Susan Dodds, and Robert Pargetter (1990 ) regard unorthodox episodes of these kinds as support for the idea that what is essential to akratic action is the presence of a second-order desire that either loses or wins against a first-order desire in the determination of action, independently of what (if anything) the agent judges it best to do. 12 For argumentation to the contrary, see Mele 1995 , chap. 4. I suggested that an agent who acts akratically against a rationally formed decisive judgment acts in a subjectively irrational way. But suppose his judgment is irrational or formed on the basis of reflection that does not take into account important, relevant attitudes of his. And suppose he acts on the basis of attitudes that constitute or reflect better reasons than the ones that ground his judgment. In that case, some have argued, his akratic action is rational, or less irrational than an action in accordance with his decisive judgment would have been (Arpaly 2000 , Audi 1990 , McIntyre 1990 ). In such cases, an akratic action may reflect an agent's system of values better than his decisive judgment does. 13 However, the fact that a certain akratic action against a particular rationally formed decisive judgment is more coherent with the agent's system of values or reasons than an action in accordance with that judgment would be is consistent with the akratic action's being subjectively irrational, to some degree, in virtue of failing to cohere with the judgment. (The same may be said of akratic actions against decisive judgments that are irrationally formed.) Also, it may be doubted that the problem with a rationally formed decisive judgment that is not sensitive to certain of the agent's values or reasons is irrationality.
4. Motivated Irrational Belief: Agency and Anti-agency Views “Biased” or “irrational” beliefs are biased or irrational relative to some standard or other. In the context of akratic belief (Davidson 1985 ; Heil 1984 ; Mele 1987 , chap. 8; Pears 1984 , chap. 4; Rorty 1983 ), the germane standards are the believer's own. In work on self-deception, general epistemic standards are typically assumed. One test for motivationally biased belief of a sort appropriate to self-deception is the following. If Sis self-deceived in believing that p,and Dis the collection of relevant data readily available to S,then if Dwere made readily available to S's impartial cognitive peers (including merely hypothetical people) and they were to engage in at least as much reflection on the issue as Sdoes and at least a moderate amount of reflection, those who conclude that pis false would significantly outnumber those who conclude that pis true (cf. Mele 2001 , 106). Two plausible requirements for impartiality in this context are that one neither
desire that pnor desire that pand that one not prefer avoidance of either of the following errors over the other: falsely believing that pand falsely believing that p. The question whether all motivationally biased beliefs are irrational raises an intriguing question about rationality. Might a motivationally biased belief be rational—or, at least, not irrational—from some legitimate point of view? W. K. Clifford asserts that “it is wrong always, everywhere, and for anyone, to believe anything upon insufficient evidence” (1886 , 346). William James denies this, contending that “a rule of thinking which would absolutely prevent me from acknowledging certain kinds of truth if those kinds of truth were really there, would be an irrational rule” ([1897 ] 1979 , 31–32). It is at least consistent with James's claim that a person who is rightly convinced that he would be miserable if he were no longer to believe that God exists may rationally believe, on insufficient evidence, that God exists. I set this question about rationality aside and pursue the issue of motivationally biased beliefs. These beliefs are irrational by general epistemic standards or from an epistemic point of view. Consider two bold theses about motivationally biased beliefs: 1.The agency view.In every instance of motivationally biased belief that p,we try to bring it about that we acquire or retain the belief that por try to make it easier for ourselves to acquire or retain it. 2.The anti-agency view.In no instance of motivationally biased belief that pdoes one try to bring it about that one acquires or retains the belief or try to make it easier for oneself to acquire or retain it. Probably, the truth lies somewhere between these poles. But which thesis is closer to the truth? One problem for the agency view is central to a familiar puzzle about self-deception. The attempts to which the view appeals threaten to undermine themselves. If Al is trying to bring it about that he believes that he is a good day trader—not by improving his investment skills, but by ignoring or downplaying evidence that he is an inferior trader while searching for evidence that he has superior investment skills—won't he see that the “grounds” for belief that he arrives at in this way are illegitimate? And won't he consequently fail in his at end p.247 tempt? A predictable reply is that the “tryings” or efforts at issue are not conscious efforts and therefore need not block their own success in this way. 14 Whether, and to what extent, we should postulate unconscious tryings in attempting to explain motivationally biased belief depends on what the alternatives are. The main problem for the anti-agency view also is linked to this puzzle about selfdeception. Apparently, we encounter difficulties in trying to understand how motivationally biased beliefs—or many such beliefs—can arise, if not through efforts of the kind the agency view postulates. How, for example, can Al's wanting it to be the case that he is a good day trader motivate him to believe that he is good at this except by
motivating him to try to bring it about that he believes this or to try to make it easier for himself to believe this? 15 The anti-agency view is faced with a clear challenge: to provide an alternative account of the mechanism(s) by which desires lead to motivationally biased beliefs. The following remarks by David Pears and Donald Davidson on the self-deceptive acquisition of a motivationally biased belief are concise expressions of two different “agency” views of the phenomenon: [There is a] sub-systembuilt around the nucleus of the wish for the irrational belief and it is organized like a person. Although it is a separate centre of agency within the whole person, it is, from its own point of view, entirely rational. It wants the main system to form the irrational belief and it is aware that it will not form it, if the cautionary belief [i.e., the belief that it would be irrational to form the desired belief] is allowed to intervene. So with perfect rationality it stops its intervention. (Pears 1984 , 87) His practical reasoning is straightforward. Other things being equal, it is better to avoid pain; believing he will fail the exam is painful; therefore (other things being equal) it is better to avoid believing he will fail the exam. Since it is a condition of his problem that he take the exam, this means it would be better to believe he will pass. He does things to promote this belief. (Davidson 1985b , 145–46) Both views rest mainly on the thought that the only, or best, way to account for certain data is to hold that the person, or some center of agency within her, tries to bring it about that she, or some “system” in her, holds a certain belief. Consider a case of self-deception similar to the one Davidson diagnoses in the passage last quoted. Carlos “has good reason to believe” that he will fail his driver's test (Davidson 1985b , 145). “He has failedtwiceand his instructor has said discouraging things. On the other hand, he knows the examiner personally, and he has faith in his own charm.The thought of failingagain is painful” (145–46). Suppose the overwhelming majority of Carlos's impartial cognitive peers presented with his evidence would believe that Carlos will fail and none would believe that he will pass. (Some peers with high standards for belief may withhold belief.) Even so, in the face of the contrary evidence, Carlos believes that he will pass. Predictably, he fails. end p.248 Self-deception is often thought to be such that if Carlos is self-deceived in believing that he will pass the test, he believed at some time that he would fail it (Bach 1981 , Demos 1960 , Haight 1980 , Quattrone and Tversky 1984 , Rorty 1972 , Sackeim and Gur 1985 ). However, in accommodating the data offered about Carlos, there is no evident need to suppose he had this true belief. Perhaps his self-deception is such that not only does he acquire the belief that he will pass, but he never acquires the belief that he will fail. In fact, this seems true of much self-deception. Seemingly, some parents who are selfdeceived in believing that their children have never experimented with drugs and some people who are self-deceived in believing that their spouses have not had affairs have never believed that these things have happened. Owing to self-deception, they have not come to believe the truth, and perhaps they never will. That having been said, there probably are cases in which a person who once believed an unpleasant truth, p, later is self-deceived in believing that p. For example, a mother who
once believed that her son was using drugs subsequently comes to believe that he has never used drugs and is self-deceived in so believing. Does a change of mind of this sort requirean exercise of agency of the kind postulated by Pears or Davidson? Is such a change of mind most plausibly explained, at least, on the hypothesis that some such exercise of agency occurred? A theorist who attends to the stark descriptions Pears and Davidson offer of the place of agency in self-deception should at least wonder whether things are so straightforward. It is often held that, in Jeffrey Foss's words, “desires have no explanatory force without associated beliefs” that identify (apparent) means to the desires' satisfaction, and this is part of “thelogic of belief-desire explanation” (1997 , 112). This claim fares poorly in the case of motivationally biased belief. “A survey of one million high school seniors found that25% thought they were in the top 1%” in ability to get along with others (Gilovich 1991 , 77). A likely hypothesis about this striking figure includes the idea that desires that pcan contribute to biased beliefs that p. If Foss's claim were true, a student's wanting it to be true that she is exceptionally affable would help explain her believing that she is only in combination with an instrumental belief (or a collection thereof) that links her believing that she is superior in this sphere to the satisfaction of her desire to be superior. But we search in vain for instrumental beliefs that are plausibly regarded as turning the trick frequently enough to accommodate the data. Perhaps believing that one is exceptionally affable can help bring it about that one is superior in this sphere, and some high school students may believe that this is so. But it is highly unlikely that most people who have a motivationally biased belief that they are exceptionally affable have that belief, in part, becausethey want it to be true that they are superior in this area andbelieve that believing that they are superior can make it so. No other instrumental beliefs look more promising. Should we infer, then, that wanting it to be true that one has a superior end p.249 ability to get along with others plays a role in explaining only relatively few unwarranted beliefs that one is superior in this area? Not at all. There is powerful empirical evidence, some of which is reviewed shortly, that desiring that pmakes a broad causal contribution to the acquisition and retention of unwarranted beliefs that p. Desires that do this properly enter into causal explanationsof the pertinent biased beliefs. It is a mistake to assume that the role characteristic of desires in explaining intentional actions is the only explanatory role they can have.
5. Evidence for and Sources of Motivationally Biased Belief If Pears or Davidson is right about a case like the mother's or Carlos's, presumably similar exercises of agency are at work in an enormous number of high school students who believe that, regarding ability to get along with others, they are “in the top 1%.” Perhaps self-deception is very common, but the same is unlikely to be true of intentional self-manipulation of the kind Pears or Davidson describes. Theorists inclined to agree
with Foss's claim about the explanatory force of desires will be inclined toward some version or other of the agency view of motivationally biased belief and self-deception. However, as I will explain, desires contribute to the production of motivationally biased beliefs, including beliefs that one is self-deceived in holding, in a variety of relatively well understood ways that fit the anti-agency model. The survey I mentioned also found that 70 percent of high school seniors “thought they were above average in leadership ability, and only 2% thought they were below average.” “A survey of university professors found that 94% thought they were better at their jobs than their average colleague” (Gilovich 1991 , 77). Data such as these suggest that desires sometimes bias beliefs. The aggregated self-assessments are wildly out of line with the facts (e.g., only 1 percent can be in the top 1 percent), and the qualities asked about are desirable. There is powerful evidence that people have a tendency to believe propositions that they want to be true even when an impartial investigation of readily available data would indicate that they probably are false. A plausible hypothesis about that tendency is that desire sometimes biases belief. Controlled studies provide confirmation for this hypothesis. In one study, 75 women and 86 men read an article asserting that “women were endangered by caffeine and were strongly advised to avoid caffeine in any form”; that the major danger was fibrocystic disease, “associated in its advanced stages with breast can end p.250 cer”; and that “caffeine induced the disease by increasing the concentration of a substance called cAMP in the breast” (Kunda 1987 , 642). (Because the article did not directly threaten men, they were used as a control group.) Subjects were then asked to indicate, among other things, “how convinced they were of the connection between caffeine and fibrocystic disease and of the connection between caffeine andcAMP on a 6point scale” (643–44). Female “heavy consumers” of caffeine were significantly less convinced of the connections than female “low consumers.” The males were considerably more convinced than the female “heavy consumers”; and there was a much smaller difference in conviction between “heavy” and “low” male caffeine consumers (the heavy consumers were slightly moreconvinced of the connections). Because all subjects were exposed to the same information and the female “heavy consumers” were the most seriously threatened by it, a plausible hypothesis is that a desire that their coffee drinking has not significantly endangered their health helps to account for their lower level of conviction (Kunda 1987 , 644). Indeed, in a study in which the reported hazards of caffeine use were relatively modest, “female heavy consumers were no less convinced by the evidence than were female low consumers.” Along with the lesser threat, there is less motivation for skepticism about the evidence. Attention to some phenomena that have been argued to be sources of unmotivatedbiased belief sheds light on motivationally biased belief. A number of such sources have been identified, including the following two.
1.Vividness of information. A datum's vividness for us often is a function of such things as its concreteness, its “imagery-provoking” power, and its sensory, temporal, or spatial proximity (Nisbett and Ross 1980 , 45). Vivid data are more likely to be recognized, attended to, and recalled than pallid data. Consequently, vivid data tend to have a disproportional influence on the formation and retention of beliefs. 2.The confirmation bias. People testing a hypothesis tend to search (in memory and the world) more often for confirming than for disconfirming instances and to recognize the former more readily (Baron 1988 , 259–65; Klayman and Ha 1987 ; Nisbett and Ross 1980 , 181–82). This is true even when the hypothesis is only a tentative one (as opposed, e.g., to a belief one has). People also tend to interpret relatively neutral data as supporting a hypothesis they are testing (Trope, Gervey, and Liberman 1997 , 115). Although sources of biased belief apparently can function independently of motivation, they also may be triggered and sustained by desires in the production of motivationallybiased beliefs. 16 For example, desires can enhance the vividness or salience of data. Data that count in favor of the truth of a proposition that one hopes is true may be rendered more vivid or salient by one's recognition that they so count. Similarly, desires can influence which hypotheses occur to one and affect the salience of available hypotheses, thereby setting the stage for the confirmation bias. 17 Owing to a desire that p, one may test the hypothesis that pis true rather than the contrary hypothesis. In these ways and others, a desire that pmay help explain the acquisition of an unwarranted belief that p. Sometimes we generate our own hypotheses, and sometimes others suggest hypotheses to us—including extremely unpleasant ones. If we were always to concentrate primarily on confirmation in hypothesis testing, independently of what is at stake, that would indicate the presence of a cognitive tendency or disposition that uniformly operates independently of desires and that desires never play a role in influencing the proportion of attention we give to evidence for the falsity of a hypothesis. However, there is powerful evidence that the “confirmation bias” is much less rigid than this. For example, in one study (Gigerenzer and Hug 1992 ), two groups of subjects were asked to test “social-contract rules such as If someone stays overnight in the cabin, then that person must bring along a bundle of firewood'” (Friedrich 1993 , 313). The group asked to adopt “the perspective of a cabin guard monitoring compliance” showed an “extremely high frequency” of testing for disconfirmation (i.e., for visitors who stay in the cabin overnight but bring no wood). The other group, asked to “take the perspective of a visitor trying to determine” whether firewood was supplied by visitors or by a local club, displayed the common confirmation bias. 18
6. A Motivational Model of Lay Hypothesis Testing An interesting recent theory of lay hypothesis testing is designed, in part, to accommodate data of the sort I have been describing. I explored it in Mele 2001 , where I offered grounds for caution and moderation and argued that a qualified version is
plausible. 19 I named it the “FTL theory,” after the authors of the two essays on which I primarily drew, Friedrich 1993 and Trope and Liberman 1996 . Here, I offer a thumbnail sketch. The basic idea of the FTL theory is that a concern to minimize costly errors drives lay hypothesis testing. The errorson which the theory focuses are false beliefs. The costof a false belief is the cost, including missed opportunities for gains, that it would be reasonable for the person to expect the belief—if false—to have, given his desires and beliefs, if he were to have expectations about such things. A central element of the FTL theory is a “confidence threshold”—or a “threshold,” for short. The lower the threshold, the thinner the evidence sufficient end p.252 for reaching it. Two thresholds are relevant to each hypothesis: “The acceptance threshold is the minimum confidence in the truth of a hypothesis,” p, sufficient for acquiring a belief that p“rather than continuing to test [the hypothesis], and the rejection threshold is the minimum confidence in the untruth of a hypothesis,” p, sufficient for acquiring a belief that p“and discontinuing the test” (Trope and Liberman 1996 , 253). The two thresholds often are not equally demanding, and acceptance and rejection thresholds respectively depend “primarily” on “the cost of false acceptance relative to the cost of information” and “the cost of false rejection relative to the cost of information.” The “cost of information” is simply the “resources and effort” required for gathering and processing “hypothesis-relevant information” (252). Confidence thresholds are determined by the strength of aversions to specific costly errors together with information costs. Setting aside the latter, the stronger one's aversion to falsely believing that p, the higher one's threshold for belief that p. These aversions influence belief in a pair of related ways. First, because, other things being equal, lower thresholds are easier to reach than higher ones, belief that pis a more likely outcome than belief that p, other things being equal, in a hypothesis tester who has a higher acceptance threshold for pthan for p. Second, the aversions influence howwe test hypotheses, not just when we stoptesting them (owing to our having reached a relevant threshold). Recall the study in which subjects asked to adopt “the perspective of a cabin guard” showed an “extremely high frequency” of testing for disconfirmation, whereas subjects asked to “take the perspective of a visitor” showed the common confirmation bias. It might be claimed that if aversions to specific errors function in the second way just identified, they work together with beliefs to the effect that testing-behavior of a particular kind is conducive to avoiding these errors. It might be claimed, accordingly, that the pertinent testing-behavior is performed with the intention of avoiding, or of trying to avoid, the pertinent error. The thrust of these claims is that the FTL theory accommodates the confirmation bias, for example, by invoking a model of intentional action. This is not a feature of the FTL model, as its proponents understand it. Friedrich, for example, claims that desires to avoid specific errors can trigger and sustain “automatic test strategies” (313), which supposedly happens in roughly the nonintentional way in which a desire that presults in the enhanced vividness of evidence for p. In Mele 2001
(41–49, 61–67), I argued that a person's being more strongly averse to falsely believing that pthan to falsely believing that pmay have the effect that he primarily seeks evidence for p, is more attentive to such evidence than to evidence that p, and interprets relatively neutral data as supporting p, without this effect's being mediated by a belief that such behavior is conducive to avoiding the former error. The stronger aversion may simply frame the topic in such a way as to trigger and sustain these manifestations of the confirmation bias without the assistance of a belief that behavior of this kind is end p.253 a means of avoiding particular errors. Similarly, having a stronger aversion that runs in the opposite direction may result in a skeptical approach to hypothesis testing that in no way depends on a belief to the effect that an approach of this kind will increase the probability of avoiding the costlier error. Given the aversion, skeptical testing is predictable independently of the agent's believing that a particular testing style will decrease the probability of making a certain error. The FTL theory applies straightforwardly to both “straight” and “twisted” self-deception (Mele 2001 , 4–5, 94–118). In straight cases, we are self-deceived in believing something that we want to be true. In twisted cases, we are self-deceived in believing something that we want to be false(and do not also want to be true). Twisted self-deception may be exemplified by an insecure, jealous husband who believes that his wife is having an affair despite possessing only relatively flimsy evidence for that proposition and despite unambivalently wanting it to be false that she is so engaged. 20 Friedrich writes: “A prime candidate for primary error of concern is believing as true something that leads [one] to mistakenly criticize [oneself] or lower [one's] self-esteem. Such costs are generally highly salient and are paid for immediately in terms of psychological discomfort. When there are few costs associated with errors of self-deception (incorrectly preserving or enhancing one's self-image), mistakenly revising one's self-image downward or failing to boost it appropriately should be the focal error” (314). Here, he plainly has straight selfdeception in mind. Whereas, for many people, it may be more important to avoid acquiring the false belief that their spouses are having affairs than to avoid acquiring the false belief that they are not so engaged, the converse may well be true of some insecure, jealous people. The belief that one's spouse is unfaithful tends to cause significant psychological discomfort. Even so, avoiding falsely believing that their spouses are faithful may be so important to some people that they test relevant hypotheses in ways that, other things being equal, are less likely to lead to a false belief in their spouses' fidelity than to a false belief in their spouses' infidelity. Furthermore, data suggestive of infidelity may be especially salient for these people and contrary data quite pallid by comparison. Don Sharpsteen and Lee Kirkpatrick observe that “the jealousy complex”—that is, “the thoughts, feelings, and behavior typically associated with jealousy episodes”—is interpretable as a mechanism “for maintaining close relationships” and appears to be “triggered by separation, or the threat of separation, from attachment figures” (1997 , 627). It certainly is conceivable that, given a certain psychological profile, a strong desire to maintain one's relationship with one's spouse plays a role in rendering the potential error of falsely believing one's spouse to be innocent of infidelity a “costly” error, in the FTL sense, and more costly
than the error of falsely believing one's spouse to be guilty. After all, the former error may reduce the probability that one takes steps to protect the relationship against an intruder. The FTL theory end p.254 provides a basis for a plausible account of twisted self-deception (Mele 2001 , chap. 5). 7. Conclusion: The Paradox of Irrationality Donald Davidson writes: “The underlying paradox of irrationality, from which no theory can entirely escape, is this: if we explain it too well, we turn it into a concealed form of rationality; while if we assign incoherence too glibly, we merely compromise our ability to diagnose irrationality by withdrawing the background of rationality needed to justify any diagnosis at all” (1982 , 303). The explanations sketched here of strict akratic action and motivationally biased belief avoid Davidson's worries about paradox. Akratic agents act for reasons, and in central cases, they make rational decisive judgments: “the background of rationality” required for that is in place. But insofar as their uncompelled actions are at odds with their rational decisive judgments, they act irrationally. Motivationally biased believers test hypotheses and believe on the basis of evidence. Again there is a background of rationality. But, owing to the influence of motivation, they violate general standards of epistemic rationality.
NOTES Parts of this chapter derive from Mele 1987 , 1995 , 1998 , and 2001. I am grateful to Piers Rawling for comments on a draft. 1. For a detailed account, see Mele 1995 , chaps. 1–7. 2. Assuming a middle ground between akrasia and self-control, not all akratic actions manifest akrasia. Someone who is more self-controlled than most people in a certain sphere may, in a particularly trying situation, succumb to temptation in that sphere against her better judgment. If her intentional action is uncompelled, she has acted akratically—even if her action manifests not akrasia but an associated imperfection. 3. An agent who makes such a judgment may or may not proceed to search for additional options. He may regard the best member of his currently envisioned options as “good enough.” 4. On failures of coherence, see Arpaly 2000 and Harman, chap. 3, this volume. 5. On the nature of motivational strength and the theoretical utility of the notion, see Mele 2003a , chap. 7. end p.255 6. For replies to Hare, Pugmire, and Watson, see Mele 1987 , chap. 2 and 51–55. See also Pugmire 1994 , responding to Mele 1987 , and the rejoinder in Mele 1995 , 44–54. 7. This is not to say that motivation is “built into” the judgment itself.
8. For opposition to the idea that desires vary in motivational strength, see Charlton 1988 , Gosling 1990 , and Thalberg 1985 . For a reply to the opposition, see Mele 2003a , chap. 7. 9. See Mele 1992 , chap. 5, for an analysis of irresistible desire. 10. This is not a necessary condition for strict akratic action. There are Frankfurt-style cases (Frankfurt 1969 ) in which, although one A-ed akratically and without any external interference, if one had been about to resist temptation, a mind-reading demon would have prevented one from doing so (see Mele 1995 , 94–95). 11. One might claim that anyone who is more strongly motivated to B than to A will also be more strongly motivated to allow that feature of her motivational condition to persist than to change it. I rebut this claim in Mele 1987 , chap. 6, and Mele 1995 , chap. 3. 12. For similar positions on akratic action, see Schiffer 1976 and Swanton 1992 , chap. 10. Swanton contends that “in the context of weakness of will, the will should be identified with strong evaluation,” a certain kind of “evaluative second-order desire” (149). Also see Jeffrey 1974 . On an alternative view, the agent who akratically A-s believes that she should believe that it is best not to A but does not believe what she believes she should (Tenenbaum 1999 ; cf. Buss 1997 , 36). 13. Dion Scott-Kakures (1997 ) argues that akratic agents are wrong about what they have “more reason” to do. 14. See, e.g., Bermudez 2000 , Quattrone and Tversky 1984 , Sackeim 1988 , and Talbott 1995 . A related response is mental partitioning: the deceived part of the mind is unaware of what the deceiving part is up to. See Pears 1984 (cf. 1991 ) for a detailed response of this kind and Davidson 1985 (cf. 1982) for more modest partitioning. For criticism of partitioning views of self-deception, see Barnes 1997 , Johnston 1988 , and Mele 1987 . 15. Two uses of “motivate” should be distinguished. In one, a desire's motivating an action or a belief is a matter of its being motivation for it. Piers's desire to fish today is motivation for him to fish, even if, desiring more to work on his chapter, he foregoes a fishing trip. In another use, a desire motivates something only if, in addition to being motivation for it, it plays a role in producing that thing. Here, I use “motivate” in the second sense. 16. I develop this idea in Mele 1987 , chap. 10, and Mele 2001 . Kunda 1990 develops the same theme, concentrating on evidence that motivation sometimes primes the confirmation bias. Also see Kunda 1999 , chap. 6. 17. For motivational interpretations of the confirmation bias, see Friedrich 1993 and Trope and Liberman 1996 , 252–65. 18. For further discussion, see Samuels and Stich, chap. 15, this volume. 19. See Mele 2001 , 31–49, 63–70, 90–91, 96–98, 112–18. 20. On this case, see Barnes 1997 , chap. 3; Lazar 1999 , 274–77; and Pears 1984 , 42– 44. Also see Davidson 1985 , 144; Demos 1960 , 589; McLaughlin 1988 , 40; Mele 1987 , 114–18; and Mele 2001 , chap. 5.
chapter 14 PARADOXES OF RATIONALITY Roy Sorensen
Fear my dot. That's right, the little round mark I used as a period. Having trouble? Would a free copy of this book help?
1. The Scope of Reasons The book bonus gives you a reason to fear the little dot. But this reason does not make fear of my dot rational. The reason is at the wrong level. You would rationally fear my dot if physicists had just persuaded you it was a black hole. A reason can make the fear rational only by conforming to the inner logic of fear—by showing that the object is dangerous. A reason to fear that the dot will harm you need not be a reason to fear that the dot will harm you (Sorensen 1998 ). In Julius Caesar, Shakespeare suggests that fear of death is irrational because the emotion makes one miserable without the prospect of a compensating benefit: Cowards die many times before their deaths; The valiant never taste of death but once. Of all the wonders that I yet have heard, end p.257 It seems to me most strange that men should fear; Seeing that death, a necessary end, Will come when it will come. (II.ii.37–42) Let us charitably interpret the poem as only a criticism of fear of death itself (and not fear of a premature death). We who cower before the Great Scythe should then concede that Shakespeare has given us a reason to suppress our fear of death. But is that enough to show that fear of death is irrational? Consider a man out on a ledge. He must escape a burning building by walking a long plank to a neighboring building. There is a real danger of falling, so the inner logic of fear is satisfied. But this fear is apt to cause a misstep. Accordingly, the man stifles his fear. (And his fear of the fear.) He averts his eyes from the abyss. He may even cultivate defiant anger against the fire to make his steps resolute. This anger is irrational. But out on the ledge, it is rational to cultivate a useful irrational emotion. (Fear of my dot may be like that.) And out on the ledge, it is rational to suppress a rational fear. (Fear of death may be like that.)
2. Emotions toward Fictions The audience watching Jaws shrieks. They seem to fear the giant shark even though they are high and dry. We shriekers know that the shark does not even exist. This leads many aestheticians to conclude that the audience is experiencing only make-believe fear (Walton 1990). Others
say that fear does not always require belief that is one in danger. They think that the mere thought of danger is enough for fear (Yanal 1999 ). These theorists are afraid of Colin Radford's (1975 ) conclusion that the members of the audience are inconsistent. Radford says members of the audience believe that they are in danger and believe that they are in no danger. This would explain why they do not flee the theater and yet are more reluctant to swim in the ocean. The young director of Jaws, Stephen Spielberg, had included a scene in which a boy is eaten by the shark. At a test screening, this spectacle caused a man in the audience to proceed to the restroom and vomit. Spielberg became distraught, fearing that his movie was “over the top.” But then the purged patron returned to his seat. Spielberg concluded that Jaws would be successful. Spielberg could agree with Radford that those who fear the shark in Jaws are being irrationally inconsistent. But Spielberg is still free to approve of the fear. Ditto for consumers of horror movies who rationally pay for the thrill of irrational end p.258 fears. Those who insist on consistency must watch Jaws under the deflating reminder that “it is only a movie.”
3. The Range of the Paradoxes Consistency is central to theoretical rationality. When I saw Jaws, I believed that the shark's first victim was a beautiful woman. I believed the woman died on July 1 because the sheriff reports she was found on July 2 and the autopsy indicated she had been dead one day. I believed a boy died shortly after, on June 29, which is the date listed on a reward poster. When I was describing the plot of Jaws, someone pointed out my inconsistent chronology. Argh! Red-faced, I withdrew my beliefs about the dates of the shark attacks. The pain of contradiction also accounts for the power of paradoxes to change our minds. Promoters of a paradox present a small set of propositions that are individually plausible but jointly inconsistent. The importance of the size of the set became apparent with the discovery of the paradox of the preface (Makinson 1965 ), an author who apologizes for the errors that are bound to be in his text. This belief expressed in the preface is correct if and only if one of the beliefs he expresses in the text is mistaken. Therefore, the modest author is bound to have a false belief. Since it is impossible for all of his beliefs to be true, they are jointly inconsistent. Does this show that it is rationally permissible to have inconsistent beliefs? Well, let's not understate the implication. Since acknowledgment of one's general fallibility seems mandatory, we wonder whether all rational human beings know themselves to be inconsistent. Practical rationality is broader than theoretical rationality. It encompasses action (as opposed to mere behavior like hiccups). Action is based on desire as well as belief, so anomalous features of preferences will affect practical rationality.
Just as causation is the cement of the physical universe, rationality is the cement of our mental lives. Causation binds each event to preceding events and future events. The result is a conglomerate of events that unifies the subject matter of physics. Rationality binds beliefs and desires into explanations of actions. The actions of a single agent fit together into a coherent whole. Agents are the subject matter of the social scientists and the humanities—and the focus of attention for any healthy human being. We all have anecdotes illustrating how the postulation of beliefs and desires creates satisfying explanations. As a newcomer to New Hampshire winters, I was puzzled to see cars parked with their windshield wipers pulled away from their windshields in apparent homage to the sky. When I started my car, I found that end p.259 my wipers were frozen to the windshield. Mystery solved! Now I too point my windshield wipers to the sky when I believe that there will be freezing rain because I want my windshield wipers mobile. The social sciences are a continuation of this commonsense practice of rationally reconstructing actions. I ask the economist: Why do so many Mexicans immigrate? He answers: Because the difference in the standard of living between Mexico and the United States is large and the cost of crossing the border is small. Thousands of Mexicans want a higher standard of living and believe moving to the United States is the most efficient means of satisfying that desire. So Mexicans come in proportion to the gap in the standard of living minus the cost of border crossing. Belief-desire explanations are also employed by biologists. Why do cheetahs hunt at noon? It is difficult to surprise prey in broad daylight. Isn't it easier and more comfortable to hunt sleeping animals in the cool of the night, under the cover of darkness? Answer: Yes, that's why lions hunt at night. Lions chase cheetahs and steal their meals. Cheetahs want to eat what they kill and believe they are more likely to be unmolested while the lions siesta. Just as folk physics contains factual errors, gaps, and conflicts, folk psychology is marred by myths, incompleteness, and inconsistency. Many paradoxes of rationality are signs of these flaws. In this respect, they are bad news. But awareness of a problem is the first step toward solving it. Optimists welcome the discovery of a paradox as an opportunity to make an improvement over raw common sense. The resulting theory itself becomes the basis of new expectations. Since the theorist's esoteric expectations can themselves be confounded, paradoxes are part of a generic cycle of self-correction.
4. Riddles of Maximization Many paradoxes orbit the principle of maximizing expected utility. Blaise Pascal formulated this principle in his seminal studies of gambling. Rational agents consider both the value of the outcome and the probability of that outcome. Since probabilities can be measured on a scale between 0 and 1, we can multiply probabilities with values to obtain expected values.
Pascal recognized that value can be measured in various ways. The egoist cares only for his well-being, the altruist cares for all. The hedonist Jeremy Bentham was later to assert that moral rightness was merely a matter of maximizing pleasure and minimizing pain. Thus Pascal's principle of maximizing expected value came to be incorporated into utilitarianism. end p.260 Pascal thought his theory of gambling afforded a new basis for theism. Whereas most arguments for believing in God's existence try to prove God exists, Pascal tries to prove only that it is prudent to believe that God exists. If God sends believers to heaven and nonbelievers to hell, then there is an incentive to believe that God exists. The payoff seems infinite. Consequently, any nonzero probability that God exists implies that the expected value of believing in God is infinite. Thus Pascal hoped to bypass epistemological limits and metaphysical deadlocks. Pascal's Wager has not filled the pews with prudent gamblers. However, it has provoked discussion about how to model infinite value. These speculations have not won the confidence of economists. They concentrate on circ*mstances involving finite values. Some use the axioms pitched to finite values as premises for the conclusion that value must be finite. This evaluative finitism is less popular with philosophers. Indeed, in the last twenty years there has been a spate of new puzzles about infinite value and some mathematically literate efforts to solve them (Vallentyne and Kagan 1997 ). Some infinite decision puzzles involve finite goods and infinite time. Before you gleams a bottle of Ever Better wine (a gift from John Pollock [1983 ]). The wine slowly improves with age. (Assume it asymptotically approaches the quality of a Château HautBrion.) More good news: You are immortal. Consequently, you are indifferent as to when you consume a particular good. When should you drink the wine? Not now. The wine will be better later. Not later. For at any given time it will be true that the wine will be even better if you waited longer. But if you do not drink the wine now and do not drink it later, then you will not drink it at all! What went wrong? You have lived according to the principle that you should not do what you will regret. Whenever you drink the wine as well, you will regret drinking it prematurely. And if you fail to drink the wine, that would be the most regrettable of all your possible alternatives. Should you abandon the prohibition against doing what you will regret? Michael Slote (1989 ) thinks Ever Better wine is a counterexample to the principle that one ought not to wittingly choose an inferior alternative. Optimizers say we should choose what is best. But when there is no best alternative, we must resign ourselves to choosing what is good enough. Slote endorses Herbert Simon's (1983 ) conclusion that rationality is about satisficing rather than optimizing. People have limited resources and so are designed to make decisions that are merely “good enough.” But rationality demands opportunism. Imagine that you are well off but could double your fortune merely by lifting a finger. Is it rationally permissible to forego lifting a finger?
How would Mother Nature answer? “Living things are not selected for their capacity to simply stay alive; they are staying alive in competition with other such living things. The trouble with satisficing as a concept is that it leaves out the competitive element which is fundamental to all life” (Dawkins 1982 , 45–46). Natural selection is sensitive to research costs. So Mother Nature will outfit us with quick and dirty heuristics that make us look like complacent satisficers. But out in the real world, one-upmanship is relentless. If not an optimizer, Mother Nature is a meliorist—always lashing her creatures on to do better. Human beings have acquired metacognitive capacities that enable us to assess how well we are doing. Human beings ask: Will I remember this or should I write it down? Is it likely that I got all these calculations correct or should I double-check? Should I abandon interviewing and just hire on the basis of résumés? When our reasoning seems to go awry, we are apt to detect the malfunction and probe for a fallacy. However, these selfdiagnoses are themselves imperfect. Often we do not know whether to blame our intuitions or our principles. Many paradoxes are rooted in this recalcitrant indecision. Constrained maximizers think the solution to the problem of when to drink Ever Better wine is a resolution (McClennen 1990 ). If you make a resolute choice to drink in ten years, the wine will be better and you will have actually drunk it. Suppose the ten years have now past. Should you keep your resolution? The unconstrained maximizer answers no: The wine will be better if you wait. The mere fact that one is breaking a resolution is irrelevant. You cannot make something a reason merely by declaring it is a reason (Broome 2001a ). Consequently, the unconstrained maximizer says that rational agent is not governed by the dead hand of the past. Only the future matters! The unconstrained maximizer concedes that this forward-looking feature of rationality is often inconvenient. When in a foreign land with little money, a sleepy traveler would like to stay in a hotel and mail the greedy proprietor her exorbitant fee upon returning to his own country. The gouging proprietor would like that as well. But both know that the traveler would have no reason to mail the exorbitant fee once he returned home. Thus the traveler must spend the night on a park bench. Punishment raises the same issue. What do you do when your threat fails to deter? Carrying through on the threat is costly to both the punisher and the punished. This cost can be justified when benefits are anticipated from maintaining the credibility of future threats. But sometimes the threatened harm outweighs the value of a steadfast reputation. If the United States had launched a massive first strike against the Soviet Union, then the Soviets would have been without a motive to carry out their threat to retaliate. And vice versa. So how did the policy of mutual assured destruction work? Why was “If you nuke us, I'll nuke you” any more persuasive than the traveler's offer to sleep now and pay later? If the president of the United end p.262 States knew that he would not counterattack, then he could not intend a counterattack. If the Soviets were to learn that the president was bluffing, then the security of the United States would be jeopardized. (The Soviets were good at detecting bluffers.) Would the president be morally obliged to cultivate a conditional intention to inflict a pointless massacre (Kavka 1978 )?
5. Good News or Causal Impact? Paradoxes have recently stimulated revisions in established theories of rationality. The most famous was precipitated by William Newcomb's problem (Nozick 1970 ) (table 14.1 ). You are offered a transparent box and an opaque box. You are free to take either or both. The opaque box might contain a million dollars. It all depends on what a talented psychologist has predicted two weeks ago. If he predicted that you would take both boxes, then he put nothing in the opaque box. If he predicted that you would take only the opaque box, then he put a million dollars in the opaque box. The psychologist is 90 percent accurate. One box or two?
Table 14.1 Empty opaque boxFull opaque box Take both boxes$1000 $1,001,000 Take one box $0 $1,000,000 On the one hand, taking two boxes dominates taking one box: whatever the contents of the opaque box, you are better off taking both boxes. On the other hand, expected utility calculations suggest that one ought take only one box: (.9 x $1,000,000) > [(.1 x $1,000,000) + $1000] The first edition of Richard Jeffrey's The Logic of Decision (1965 ) implies that the onebox decision is correct. His book articulates an evidentialist theory that counsels us to maximize good news. Anyone faced with Newcomb's choice should hope that he finds himself walking away with just one box, for the expected value of that box is $900,000. If you find yourself exiting with both boxes, you will be disappointed to learn that the total expected value is only $101,000. One-boxers and two-boxers agree that you are better off being a one-boxer in a Newcomb situation. But two-boxers believe that this is because Newcomb scenarios reward irrational individuals. When fools are rewarded and sages are penalized, it is best to be a fool. However, this is a truth that can never be practically applied. Weakness of will aside, any action I choose is viewed by me as the best of my available options (even when this option is “looking like a fool”). Causal decision theorists are two-boxers. They say you should behave in a end p.263 way that causes the best outcome. Producing and sustaining benefits in the world is what counts—not cheering yourself up with statistics. Since Richard Jeffrey found the two-box solution compelling, he retracted the actuarial aspects of his classic text. His second edition (Jeffrey 1983 ) accommodates causal decision theory.
Presently, the vast majority of philosophers are two-boxers. Hence most commentators think Newcomb's problem stimulated a fundamental improvement of decision theory.
6. Winning with Irrationality Jeffrey's revision also affects game theory. In game theory, the outcome depends on your choice plus choices of other people (not just states of nature). For instance, California teenagers play “chicken.” They drive cars toward each other until one of them, the chicken, swerves to avoid a head-on collision. Losers in these games of brinkmanship are chagrined by how they are disadvantaged by their rationality. If persuaded that the other players are irrationally undeterred by the looming mutual disaster, capitulation is rationally mandatory. Little wonder that Nikita Khrushchev would feign irrationality to prevail in Cold War confrontations with the United States. In his classic The Strategy of Conflict (1960), Thomas Schelling described a variety of realistic circ*mstances in which it pays to lose control, to block your acquisition of information, and to engage in other practices that fit the stereotype of irrationality. But these are all cases in which the irrationality is only apparent. Evolutionary explanations of rationality generally assume that rationality arose because it leads to reproductive success. Cases in which irrationality really is advantageous force qualifications to this account. end p.264
7. Collective Irrationality from Individual Rationality The Prisoners' Dilemma (see Bicchieri, chap. 10, and Dreier, chap. 9, this volume) was discovered in 1950 by Merrill Flood and Melvin Dresher while employed by the RAND Corporation. Albert W. Tucker illustrated the game with a tale of two prisoners. The prosecutor has enough evidence to lightly penalize each prisoner. She separately offers leniency to each prisoner in exchange for a confession that could be used against his accomplice. Specifically, she presents row and column with the payoff matrix in table 14.2 . Table 14.2 Defect Do not defect Defect 3 years, 3 years1 year, 4 years Do not defect4 years, 1 year 2 years, 2 years As in Newcomb's problem, there is a dominant choice: defecting leaves each prisoner better off regardless of how the other chooses. The resemblance to Newcomb's problem can be strengthened: Suppose that each prisoner knows that his accomplice is so similar
that there is a 90 percent chance that his accomplice will make the same choice as he does. Defecting then yields a worse expected penalty (2.8 years) than not defecting (2.2 years). Game theorists are far more persuaded by the dominance argument for defecting than the expected utility argument for not defecting.
8. The Iterated Prisoners' Dilemma The case for defecting assumes, for the sake of simplicity, that the prisoners are certain they will not have further contact. If the prisoners think they can penalize each other for defecting, then those consequences must be added to the overall package of sanctions. Since people generally have the prospect of interacting in the future, there is less practical worry about betrayal. In particular, if Prisoners' Dilemma were played repeatedly, one would expect much cooperation. Flood and Dresher tested this prediction by having two friends play Prisoners' Dilemma one hundred times in sequence (Poundstone 1992 , 106–16). The players usually cooperated (one cooperated 68 times, the other 78 times). The running log each player kept of the reasoning for his move shows that each realized that there was no selfinterested motive to cooperate in the last round in the game. Accordingly, each defected on round 100. But given that each side knows that the other will defect at round 100, neither has a motive to cooperate on round 99. The only point in cooperating was to affect one's reputation in a way that would lead to a better outcome. Thus game end p.265 theory appears to predict both sides will also defect on play 99. This reasoning is also available to the players, and so they should also defect on play 98. Indeed, this chain of reasoning applies round by round, all the way back to the first play. Therefore there should be no cooperation at all! What actually happened in the experiment is that one player cooperated on play 99 and the other defected. Prior to round 99 there was mostly cooperation. Subsequent experiments have found that people cooperate on most plays of an iterated Prisoners' Dilemma. Game theorists are chagrined by how ordinary people cooperate to achieve higher payoffs than is theoretically possible. A majority of economists accept the soundness of the reasoning in the iterated Prisoners' Dilemma. They concede that in ordinary circ*mstances, iterating Prisoners' Dilemma leads to more cooperation. But that is only because the ordinary agents do not have the certainty enjoyed by the hypothetical prisoners. Ordinary agents are a bit unsure about whether the conditions of the game really hold. In general, the economists draw the moral that the assumption that the situation is “common knowledge” between the players is a surprisingly strong condition that is rarely satisfied in our ordinary dealings with one another. The Alice in Wonderland atmosphere of Prisoners' Dilemma just reflects the hidden strength of this apparently innocuous simplification.
9. Medieval Backward Inductions? Since the iterated Prisoners' Dilemma acquired notoriety, there has been a wave of backward inductions: the chain-store paradox, the centipede, and so on (see Bicchieri, chap. 10, this volume). Instead of reviewing these spin-offs, I will consider two precursors of this influential slippery-slope argument. Robert Louis Stevenson's 1893 short story “The Bottle Imp” features a work of the Devil: an indestructible bottle that contains an imp. The imp grants the bottle's owner nearly anything he wishes. The catch is that anyone who dies while still possessing the bottle goes to hell. If you try to throw the bottle away, it comes back. The only way to end possession is to sell the bottle at a cheaper price to someone who is well informed about the bottle's terms of ownership. In Stevenson's tale, the bottle starts out at an enormous price. It passes through famous hands: Napoleon, Captain Cook, etc. Eventually a boatswain agrees to buy the bottle for the price of two centimes, a centime being the lowest unit of currency in the world. When the seller dutifully points out that the buyer is sure to go to hell, the boatswain explains that he is going to hell anyway. But what if the Devil had the foresight to require that the buyer not already be damned? Richard Sharvey (1983 ) notes that if we close such loopholes, we get a “proof” that the bottle cannot be sold for any amount of money. The bottle cannot be sold if it was bought at the lowest unit of currency, say, one centime. Thus no well-informed person will buy the bottle for two centimes. And if no one will buy at two centimes, no one will buy at three centimes. And so up to any amount of money! Yet it seems that some rational and well-informed person would buy it for a million centimes. In the first edition of “The Bottle Imp,” Stevenson erroneously credits the idea to “B. Smith.” This is the garbled nickname of Richard John Smith, who was an actor in a play entitled “The Bottle Imp” (Beach 1910 ). The basic plot is in German folklore. The Brothers Grimm borrow it from Johann Jakob Christoffel von Grimmelshausen's Die Landstortzerin Courasche, published in 1756 (Starck 1911 ). This suggests that the idea originated as fable or a medieval legend. Since the plot was used in stories all over Europe, the insight behind Richard Sharvey's backward induction argument may be a thousand years old. A more immediate precursor to the iterated Prisoners' Dilemma appears to have been discovered during World War II by the Swedish mathematician Lennart Ekbom (Sorensen 1988 , 253). The “prediction paradox” was in discussion at Princeton University shortly after the war (at least among Kurt Gödel and his students), and so may have reached the ears of Flood or Dresher and thereby primed the discovery of the iterated Prisoners' Dilemma. Nowadays the prediction paradox is most frequently posed as follows: A teacher announces that there will be a surprise test next week. A clever student protests that the surprise test is impossible: “Everybody knows we meet three times next week: Monday, Wednesday, and Friday. If the test were given on Friday, then on Thursday we could foresee the day of the “surprise” test. We can therefore eliminate the possibility that the test will be on Friday. This reasoning is available to us on Tuesday night. So if a test has yet to occur by then, we would know the test is on Wednesday. Only Monday remains. Therefore, a test on Monday would not be a surprise. This chain
of reasoning would be available to us on Sunday, so a Monday test would be expected. Therefore, the announced surprise test is impossible.” end p.267 The first two articles published on the prediction paradox accepted the clever student's reasoning. But since 1950, each of the subsequent hundred or so commentators has rejected the student's reasoning. They think the argument fails even if the agents are idealized in the manner envisaged by economists. The philosophers' unanimous verdict of unsoundness contrasts resoundingly with the widespread acceptance of backward induction arguments by economists. Although there is a consensus that the clever student's reasoning is unsound, philosophers disagree about how it specifically fails. Many commentators connect the prediction paradox to G. E. Moore's problem of explaining what is odd about saying, “It is raining but I do not believe it.” Although this sentence is consistent, I cannot rationally believe it. This is odd because the sentence merely describes an error on my part. As underscored by the preface paradox, I can believe that some belief or other of mine is mistaken. But I cannot believe a specific proposition is falsely believed by me. This asymmetry of believability suggests a solution to the prediction paradox. The teacher's announcement merely guarantees that the clever student will fail to have a true belief about the test date. The student can rationally believe this announcement. However, with the accumulation of test-less dates, the announcement's implications become more specific. By Thursday night, the teacher's announcement implies, “The test is on Friday but you do not believe it.” Since the student cannot believe this specific attribution of an error of omission, he can no longer know that the announcement is true. Therefore, if the test is on Friday, he will be surprised by it.
10. Reflexive Beliefs Many people who are introduced to the surprise test paradox make the following observation: If the students are persuaded by the argument, then they will not expect a test. Therefore, if the teacher gives the test, it will be a surprise. It is self-defeating to believe that the test cannot be given. When surprise is your goal, you try to predict what your target will believe. The main basis for prediction is the principle of charity. This principle instructs us to interpret agents as rational agents. We try to make their beliefs and desires cohere. This is largely a matter of ensuring that their beliefs come out as largely true. Since we must judge truth by our own lights, this winds up as a policy of maximizing agreement. A problem arises when the goal is to disagree with those whom we are trying to interpret. I introduce my students to this discoordination dilemma with a multiple choice question: Which answer will be chosen by the fewest classmates? end p.268
(a) (b) (c) (d) (e) Students think of a reason why one answer would be the least popular. But this reasoning is undermined by the likelihood that other students would replicate the reasoning. Travelers face a less pure version of this problem when trying to avoid congestion. They try to pick unpopular times and routes but worry that “great minds think alike.” A prediction is reflexive if the probability of the prediction is affected by its conditions of dissemination. When President Bush took office in 2001, he predicted an economic downturn. People reacted in ways that helped the prediction come true. Reflexive predictions vary in how sensitive they are to subjective reactions to them. At the logical limit, and perhaps as a degenerate case, lies one of Jean Buridan's sophisms, (B) You do not believe this sentence. If you do not believe the sentence, it is true and so you should believe it. Thus you cannot have a rational attitude toward the sentence. Others can. If their observation shows that you believe the sentence, then they rationally believe that the sentence is false. If their observation shows that you do not believe the sentence, then they rationally believe the sentence is true. Buridan's sentence can be impersonalized. If a rational person believes “No rational person believes this sentence” then it is true and so merits belief. But that belief would make it false and so warrants disbelief in it. The truth-teller analogue of this paradox is also paradoxical (Cave 2001 ). If you believe “You believe this statement,” then your belief must be true. But can you believe it? If you believe it, then there is some proposition that you believe. The “this” in “You believe this statement” generates another this if we try to make the proposition the result of substituting the sentence “You believe this statement.” Since the “this” cannot be unpacked, some conclude that “You believe this statement” is meaningless. The same objection would apply to Buridan's “You do not believe this statement.” In the literature on the liar paradox, there is a precedent for this objection and a precedent for its solution (Quine 1987 , 148). Consider (B) “Does not yield a statement you believe when appended to its own quotation” does not yield a statement you believe when appended to its own quotation. end p.269 Since the thirteen-word quotation is a noun rather than a demonstrative, it does not lead to the unpacking regress. However, there still seems to no evidential basis to believe “I believe this statement.” Consider a patient who realizes he is taking a placebo. He has read that even people who realize that they are taking a placebo tend to get better in virtue of their belief that they will get better. So he believes he will recover because of his belief that he will recover. Optimism in his recovery can be sustained by a preexisting belief that he will recover.
But how does he get his original belief going? That original belief seems to be without any evidential base. Although we can view others as believing p without the benefit of evidence, we cannot picture ourselves as believing p without any evidence. “I have no evidence that God exists but I believe that God exists” has the same peculiarity as Moore's sentence “It is not raining but I believe it is raining.”
11. Preference Gaps Strangely, Jean Buridan is most famous for a paradox there is no record of him inventing. “Buridan's ass” is indifferent between two bales of hay. If a choice of x over y implies a preference for x over y, then the ass starves. Buridan's ass was a popular exhibit in commentaries on the principle of sufficient reason (which says nothing happens without a reason). If we were able to starve asses by presenting them with equally appealing goods, then “Choice implies preference” would be empirically confirmed. Since we do not find asses transfixed in fatal indecision, there seems to be formidable empirical data against the principle of sufficient reason. However, Gottfried Leibniz, the famous champion of the principle, maintained it is impossible to starve an ass with two equal goods. For Leibniz denied there could be two equal goods. Leibniz appears to grant that his adversaries sometimes believe that two apples are equally good. But he does not take this belief in the equality of two goods as sufficient for indifference between the two apples. According to Leibniz, there is always at least a slight preference. Biologists attribute less sensitivity to the ass. Arbitrary choice seems to underlie the erratic behavior of many animals (Driver and Humphries 1988 ). This is just what game theory predicts. Randomizing is a defense against predators who need to predict your hiding spot or anticipate your course of flight. This is especially true if you lack the brains to outwit predators. end p.270 Some people are reluctant to admit that animals choose arbitrarily because they associate unpredictable action with a free will. This left-handed respect for free will leads to pessimism about the social sciences. How could there be science governing people without laws determining human choice? Emile Durkheim answered that the unpredictability of individuals was compatible with the predictability of classes of people. An actuary cannot predict which individuals will commit suicide, but can predict what percentage of people will commit suicide. The aggregate behavior of an individual is also predictable in this statistical sense. The butterfly's zigzag flight divides up into so many zigs and so many zags. The relative frequency of each would aid a ballistics expert in his attempt to shoot down the butterfly. Since physicists accept irreducibly statistical predictions when studying radioactive piles, most social scientists do not think that random choice is methodologically worrisome. After all, they themselves use random-control trials. A Martian studying human
psychologists would not be able to predict whether the randomizing psychologist was going to allocate a particular subject to the treated group or to the control group.
12. Weakness of Will If I regard two incompatible alternatives as equally belief-worthy, then I cannot believe one over the other. If I regard them as equally desirable, I cannot desire one over the other. But if I regard two incompatible alternatives as equally choice-worthy, I can choose one over the other. Therefore, beliefs and desires do not determine all choices. If there is weakness of will, then there are graver gaps in belief-desire framework (see Mele, chap. 13, this volume). When a dieter sheepishly picks a brownie over a fruit salad, he insists that, all things considered, he prefers the fruit salad. Yet he chooses the brownie. Thus he seems to wittingly take the inferior alternative. Anyone who accepts weakness of will accepts a significant limit on the predictive and explanatory power of beliefs and desires. Indeed, he seems to be undermining the very point of postulating beliefs and desires. The prediction deficit could be filled with a supplementary theory that described when and how weakness of will takes place. There are many commonsense generalizations about the conditions under which people fall prey to temptation. Backsliders love company. The dieter is more apt to break his resolution if the temp tation is placed under his nose, without the notice of those who monitor whether he is on a diet, and among others who are partaking—especially when they themselves are falling to temptation. Social psychologists have elaborated such generalizations into a theory of eating. But the generalizations owe their success to a secondary effort to explain the dieter in terms of beliefs and desires—just different beliefs and desires than we first assumed. The weakness of will seems to disappear. Although existence of weak-willed acts seems to be a truism, it is difficult to make detailed sense of it. Economists have an especially strong tendency to follow the saying “Actions speak louder than words”: the dieter's choice shows that he actually prefers the brownie over the fruit salad!
13. Ill-Structured Preferences Economists are notoriously nonjudgmental about the content of preferences. They deny a priori that the ultimate goals of Osama bin Laden could be irrational. Most economists accept David Hume's principle that “reason is and ought to be the slave of the passions.” The economists share commonsense fussiness about the structure of our preferences. Almost everyone demands indifference between alternatives that are logically equivalent. People are vulnerable to framing effects in which logically equivalent alternatives are described differently. Patients and physicians are more likely to choose a treatment that is couched in terms of chances of survival rather than chances of death. These preferences
do not survive exposure. Once we realize that we prefer one alternative over a logical equivalent, we lapse into indifference. Some of our questionable preferences about risk are more robust. Most Americans know that mile for mile, they are far more likely to be injured in a car than in an airplane. Yet the greater control afforded by a car leads them to frequently choose an eight-hour car ride over a one-hour plane trip. An economist might try to construe this as a rational choice by postulating a taste for control. But those less committed to making consumers come out rational will condemn such travelers as self-destructively muddled. Other robust preferences about risk are more difficult to dismiss. The St. Petersburg Paradox features a fair coin that will be tossed until a head results. You will then be paid $2 n–1 where n equals the number of tosses. So the expected return is: (1/2 × $1) + (1/4 × $2) + (1/8 × $4) ++ (1/2n × $2 n–1 ) + end p.272 Since each addend equals a half dollar, and there are infinitely many of them, the sum is infinite. Thus someone who maximized expected money should be willing to pay any amount of money for this bet. Yet few people would pay $100 for the deal. Daniel Bernoulli (1738 ) denied that the principle of maximizing expected utility implies that the deal is of infinite value (even if infinite money were possible). He pointed out that doubling one's cash holdings from 1 million to 2 million does not really double its value to you. Each new dollar tends to have less influence on your welfare than the preceding dollar. Bernoulli's insight is enshrined in contemporary economics as the law of the diminishing marginal utility of money. The rate of diminution resists precise calculation, but Bernoulli inferred that it is a logarithmic function. This would preclude infinite sums. There is more to the psychology of risk than the diminishing marginal utility of money. Maurice Allais (1979 ) emphasizes that people also care about the dispersion of the possible payoffs (see Dreier, chap. 9, this volume). The St. Petersburg bettor has possible future selves spread out into a mass of paupers and a mass of tycoons with hardly anything in between. This aversion to wide ranging outcomes is manifested most strikingly in our taste for certainty. Daniel Kahneman and Amos Tversky (1979 ) firmed up the empirical basis for the “Allais paradox” with a number of experiments in Israel. At the time of their study, the median net monthly income for a family was 3,000 Israeli pounds. Israelis were asked for their preferences between the following options: Option A: 4000 with a probability of .8 Option B: 3000 with certainty Option C: 4000 with a probability of .2 Option D: 3000 with a probability of .25 The majority of Israelis preferred B over A even though A has a higher expected return of 3,200. This cannot be explained by the diminishing marginal utility of money, because many of the same people prefer C (expected return 800) over D (expected return 750).
People also dislike “risky” processes. Daniel Ellsberg (1961 ) has us consider two lotteries. In the first lottery, you get a $100 prize if you choose a red marble from an urn that you know to be composed of 50 red marbles and 50 black marbles. In the second lottery, the urn contains red marbles and black marbles in an unknown ratio. You get $100 if you pick a red marble. Standard decision theory says you should be indifferent between the two lotteries because they yield the same expected return. Yet a large percentage of people prefer the first lottery. Many continue to maintain the preference for the better known urn even after the equal expected return is made salient. There are important anomalies of preference that are independent of risk. Breakdowns of the transitivity of preference are the most discussed (see Dreier, end p.273 chap. 9, and Joyce, chap. 8, this volume). If a shopper prefers car A over car B and car B over car C, then he should prefer car A over car C. If the shopper instead prefers car C over car A, then the car salesman will be struck by the “inconsistency.” He might try to prove the irrationality of the shopper by converting him into a money pump. If the shopper prefers C over A, then he should be willing to trade A plus a dollar to get C. Since he prefers B over C, the shopper should trade C plus a dollar to get B. And since he prefers A over B, the shopper should trade B plus a dollar to get A. Although each of the three trades is fair (the shopper always swaps for a more preferred basket of goods), the shopper winds up in the same position (owning car A) minus three dollars. The absurdity can be magnified by repeating the cycle over and over; all the shopper's money could be drained away by the car salesman.
14. Social Choice Transitivity also breaks down for collective preferences. The marquis de Condorcet has us imagine that a group of three individuals is governed by majority rule. Their preferences are individually transitive: Voter 1: A, B, C Voter 2: B, C, A Voter 3: C, A, B Yet the group's preference is not transitive. By majority rule, the group prefers A over B (because voters 1 and 3 vote for A over B), B over C (because voters 1 and 2 vote for B over C), and C over A (because voters 2 and 3 vote for C over A). Given this cyclic preference structure, the group decision will depend on the order in which the issues are voted upon. If we start by asking whether A or B should be eliminated, the group will eliminate B. If we then continue the deliberation by asking whether A or C should be eliminated, then A will be eliminated and the group will ultimately decide in favor of C.
But had the first stage of the vote begun with the question of whether B or C should be eliminated, then C would have been eliminated and A would have ultimately prevailed. Democrats offered increasingly complex repairs to this vulnerability to sophisticated voting. Kenneth Arrow (1951 ) ended the quest for a perfect “social aggregation device” by proving that all democratic voting schemes must form some cyclical preferences. Perfect democracy is impossible—and not merely because it must be shaped with imperfect human clay. end p.274 Condorcet's insight about cyclical majorities applies to individual people insofar as they are pictured as collectives. Consider self-deception (see Mele, chap. 13, this volume). If Cicero deceives Tully into believing p, then Tully must believe. Yet it must also be the case that Cicero does not believe. After all, if Cicero believed p, then he would be trying to lead Tully to a truth, not a falsehood. These two requirements preclude the possibility that Cicero is identical to Tully. For if Cicero is Tully, then Cicero does not believe p (to qualify as a deceiver) and does believe p (to qualify as the victim of the deceiver). A common solution is to say that part of the self-deceiver believes p and part of him does not. These parts are homunculi, “little men” who are capable of deceiving each other. This homuncular model is the norm in explanations of illusions. If we are collectives of homunculi, then it is probable that some of our preferences are formed by majority rule. Hence, even individuals should be vulnerable to Condorcet's paradox. Indeed, the contemporary revival of faculty psychology suggests that any paradox of collective rationality should arise in miniature at the level of individuals.
NOTE I thank Andreas Teaber for his Stevenson scholarship.
part ii RATIONALITY IN SPECIFIC DOMAINS end p.277 end p.278
chapter 15 RATIONALITY AND PSYCHOLOGY Richard. Samuels Stephen. Stich Since the early 1970s, psychologists have devoted a great deal of attention to human reasoning and decision making, and to the psychological processes that underlie them. While some of this attention was motivated by the intrinsic interest and importance of
these processes, much of it was provoked by a series of experimental findings that, in the view of many, had “bleak implications” for human rationality (Nisbett and Borgida 1975 ). In this essay we'll begin, in section 1, by presenting a brief sketch of some of these disturbing findings, most of which were reported by psychologists in what has become known as the “heuristics and biases” tradition. In section 2, we'll set out three increasingly pessimistic interpretations of the findings that have been suggested by a number of authors. There have been many challenges to these pessimistic interpretations. One of the most interesting and influential challenges was launched, in the early 1990s, by a group of researchers working in the then newly emerging interdisciplinary field of evolutionary psychology. This challenge, and the experimental findings that support it, will be our focus in section 3. On the basis of these new findings, evolutionary psychologists have suggested a variety of much more optimistic views about the rationality of ordinary people. In section 4, we'll sketch three of these views. We are inclined to think that the right reaction to the entire body of findings on human reasoning is significantly less pessimistic than the most dire interpretation end p.279 suggested by writers in the heuristics and biases tradition, but rather more pessimistic than suggested by the Panglossian pronouncements of some evolutionary psychologists. In section 5, we'll defend this “middle way” and sketch a family of “dual processing” theories of reasoning that, we'll argue, offer some support for the moderate interpretation we advocate.
1. Some Unsettling Studies of Human Reasoning In 1966, Peter Wason published a highly influential study of a cluster of reasoning problems that became known as the selection task. By 1993, the selection task had become “the most intensively researched single problem in the history of the psychology of reasoning.” (Evans, Newstead, and Byrne 1993 , 99) Figure 15.1 illustrates a typical example of a selection task problem. What Wason and numerous other investigators have found is that subjects typically perform very poorly on questions like this. Most subjects respond correctly that the E card must be turned over, but many also insist that the 5 card must be turned over, though the 5 card could not falsify the claim no matter what is on the other side. Also, a majority of subjects maintain that the 4 card need not be turned over, though without turning it over there is no way of knowing whether it has a vowel on the other side. Subjects do not do poorly on all selection task problems, however. A wide range of variations on the basic pattern has been tried, and on some versions of the problem a much larger percentage of subjects answers correctly. These results form a bewildering pattern, since there is no obvious feature or cluster of features that separates versions on which subjects do well from those on which they do poorly. Much of the experimental literature on theoretical reasoning has focused on tasks that concern probabilistic judgment. Among the best-known experiments of this kind are
those that involve so-called conjunction problems. In one quite famous experiment, Tversky and Kahneman (1982 , p. 92) presented subjects with the following task. Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice and also participated in anti-nuclear demonstrations. end p.280
Figure 15.1 Please rank the following statements by their probability, using 1 for the most probable and 8 for the least probable. (a) Linda is a teacher in elementary school. (b) Linda works in a bookstore and takes Yoga classes. (c) Linda is active in the feminist movement. (d) Linda is a psychiatric social worker. (e) Linda is a member of the League of Women Voters. (f) Linda is a bank teller. (g) Linda is an insurance sales person. (h) Linda is a bank teller and is active in the feminist movement. In a group of naive subjects with no background in probability and statistics, 89 percent judged that (h) was more probable than (f) despite the obvious fact that one cannot be a feminist bank teller unless one is a bank teller. When the same question was presented to statistically sophisticated subjects—graduate students in the decision science program of the Stanford Business School—85 percent gave the same answer! Results of this sort, in which subjects judge that a compound event or state of affairs is more probable than one of the components of the compound, have been found repeatedly since Kahneman and Tversky's pioneering studies, and they are remarkably robust. This pattern of reasoning has been labeled the conjunction fallacy. Another well-known cluster of studies examines the way in which people use base-rate information in making probabilistic judgments. According to the familiar Bayesian
account, the probability of a hypothesis on a given body of evidence depends, in part, on the prior probability of the hypothesis. However, in a series of elegant experiments, Kahneman and Tversky (1973 ) showed that subjects often seriously undervalue the importance of prior probabilities. One of these experiments presented half of the subjects with the following “cover story.” A panel of psychologists have interviewed and administered personality tests to 30 engineers and 70 lawyers, all successful in their respective fields. On the basis of this information, thumbnail descriptions of the 30 engineers and 70 lawyers have been written. You will find on your forms five descriptions, chosen at random from the 100 available descriptions. For each description, please indicate your probability that the person described is an engineer, on a scale from 0 to 100. (p. 54) The other half of the subjects was presented with the same text, except the “base-rates” were reversed. These subjects were told that the personality tests had been administered to 70 engineers and 30 lawyers. Some of the descriptions that were provided were designed to be compatible with the subjects' stereotypes of engineers, though not with their stereotypes of lawyers. Others were designed to fit the lawyer stereotype, but not the engineer stereotype. And one was intended to be quite neutral, giving subjects no information at all that would be of use in making their decision. Here are two examples, the first intended to sound like an engineer, the second intended to sound neutral: Jack is a 45-year-old man. He is married and has four children. He is generally conservative, careful, and ambitious. He shows no interest in political and social issues and spends most of his free time on his many hobbies, which include home carpentry, sailing, and mathematical puzzles. Dick is a 30-year-old man. He is married with no children. A man of high ability and high motivation, he promises to be quite successful in his field. He is well liked by his colleagues. (p. 54) As expected, subjects in both groups thought that the probability that Jack is an engineer is quite high. Moreover, in what seems to be a clear violation of Bayesian principles, the difference in cover stories between the two groups of subjects had almost no effect at all. The neglect of base-rate information was even more striking in the case of Dick. That description was constructed to be totally uninformative with regard to Dick's profession. Thus, the only useful information that subjects had was the base-rate information provided in the cover story. But that information was entirely ignored. The median probability estimate in both groups of subjects was 50 percent. Before leaving the topic of base-rate neglect, we want to offer one further example illustrating the way in which the phenomenon might well have serious practical consequences. Here is a problem that Casscells, Schoenberger, and Gray end p.282 boys (1978 , p. 999) presented to a group of faculty, staff, and fourth-year students at Harvard Medical School. If a test to detect a disease whose prevalence is 1/1000 has a false positive rate of 5%, what is the chance that a person found to have a positive result actually has the disease, assuming that you know nothing about the person's symptoms or signs? %
Under the most plausible interpretation of the problem, the correct Bayesian answer is 2 percent. But only 18 percent of the Harvard audience gave an answer close to 2 percent. Forty-five percent of this distinguished group completely ignored the base-rate information and said that the answer was 95 percent. One of the most extensively investigated and most worrisome cluster of phenomena explored by psychologists interested in reasoning and judgment involves the degree of confidence that people have in their responses to factual questions—questions like: In each of the following pairs, which city has more inhabitants? (a) Las Vegas (b) Miami (a) Sydney (b) Melbourne (a) Hyderabad (b) Islamabad (a) Bonn (b) Heidelberg In each of the following pairs, which historical event happened first? (a) Signing of the Magna Carta (b) Birth of Mohammed (a) Death of Napoleon (b) Louisiana Purchase (a) Lincoln's assassination (b) Birth of Queen Victoria After each answer subjects are also asked: How confident are you that your answer is correct? 50% 60% 70% 80% 90% 100% In an experiment using relatively hard questions it is typical to find that for the cases in which subjects say they are 100 percent confident, only about 80 percent of their answers are correct; for cases in which they say that they are 90 percent confident, only about 70 percent of their answers are correct; and for cases in which they say that they are 80 percent confident, only about 60 percent of their answers are correct. This tendency toward overconfidence seems to be very robust. Warning subjects that people are often overconfident has no significant effect, nor does offering them money (or bottles of Champagne) as a reward for accuracy. end p.283 Moreover, the phenomenon has been demonstrated in a wide variety of subject populations including undergraduates, graduate students, physicians, and even CIA analysts. (For a survey of the literature, see Lichtenstein, Fischoff, and Phillips 1982 .) The studies we've reviewed so far have focused on subjects' normatively problematic performance on belief formation and judgmental tasks. But there is also a large experimental literature that seems to indicate that human decision making processes are normatively problematic. Since space is limited, we'll recount only one example, albeit a
particularly disturbing one. Tversky and Kahneman (1981 ) presented a group of subjects with the following problem: Imagine that the U.S. is preparing for the outbreak of an unusual Asian disease, which is expected to kill 600 people. Two alternative programs to combat the disease have been proposed. Assume that the exact scientific estimate of the consequences of the programs is as follows: If Program A is adopted, 200 people will be saved. If Program B is adopted, there is a 1/3 probability that 600 people will be saved, and a 2/3 probability that no people will be saved. A second group of subjects was given an identical problem, except that the programs were described as follows: If Program C is adopted, 400 people will die. If Program D is adopted, there is a 1/3 probability that nobody will die and a 2/3 probability that 600 people will die. On the first version of the problem a substantial majority of the subjects chose program A. But on the second version most chose program D, despite the fact that the outcome described in A is identical to the one described in C. The disconcerting implication of this study is that the decisions we make are strongly influenced by the manner in which the options are described or framed.
2. What Do These Results Show? Three Pessimistic Views What do these results and the many similar results to be found in the experimental literature 1 tell us about the rationality of ordinary people's reasoning and decision end p.284 making and about the mental mechanisms that underlie those processes? In this section we'll distinguish three answers to this question that have been suggested in the literature—answers that get increasingly pessimistic (and, as we will argue in section 5, increasingly implausible). Before attempting to the answer the question, we must, of course, adopt some normative standard or metric for assessing the rationality of inferences and decisions. Though researchers in this area rarely offer an explicit and general normative theory of rationality, we think that most authors tacitly adopt some version of what Edward Stein has called the “Standard Picture” of rationality: “According to this picture, to be rational is to reason in accordance with principles of reasoning that are based on rules of logic, probability theory and so forth. If the standard picture of reasoning is right, principles of reasoning that are based on such rules are normative principles of reasoning, namely they are the principles we ought to reason in accordance with” (Stein 1996 , 4). Thus the Standard Picture maintains that the appropriate criteria against which to evaluate human reasoning are the rules derived from formal theories such as classical logic, probability theory, and decision theory. 2 So, for example, one might derive something like the following principle of reasoning from the conjunction rule of probability theory:
Conjunction Principle: One ought not assign a lower degree of probability to the occurrence of event A than one does to the occurrence of A and some (distinct) event B. (Stein 1996 , 6) Given principles of this kind, one can evaluate the judgments and decisions of human subjects and the mechanisms that produce them. To the extent that a person's judgments and decisions accord with the principles of the Standard Picture, they are rational, and to the extent that they violate such principles, they fail to be rational. Similarly, to the extent that a reasoning or decision making mechanism produces judgments that accord with the principles of the Standard Picture, the mechanism is rational and to the extent that it fails to do so, it is not rational. If we adopt the Standard Picture, then one quite plausible conclusion to draw from the experimental findings reported in the heuristics and biases literature is that (1) People's intuitive judgments on a large number of reasoning and decision making problems regularly deviate from appropriate norms of rationality. To understand a second claim that has been made on the basis of the experimental findings, we need to recount how researchers in the heuristics and biases tradition explain people's performance on many reasoning problems. The basic end p.285 explanatory strategy is to posit the existence of reasoning “heuristics”—rules of thumb that people employ when reasoning. In the case of the conjunction fallacy and base rate neglect, for example, Kahneman and Tversky propose that people often rely on what they call the representativeness heuristic. Given specific evidence (e.g. a personality sketch), the outcomes under consideration (e.g. occupations or levels of achievement) can be ordered by the degree to which they are representative of that evidence. The thesis of this paper is that people predict by representativeness, that is, they select or order outcomes by the degree to which the outcomes represent the essential features of the evidence. In many situations, representative outcomes are indeed more likely than others. However, this is not always the case, because there are factors (e.g. prior probabilities of outcomes and the reliability of evidence) which effect the likelihood of outcomes but not their representativeness. Because these factors are ignored, intuitive predictions violate statistical rules of prediction in systematic and fundamental ways. (Kahneman and Tversky 1973 , 48) If explanations of this sort are correct, then we can conclude that: (2) Many of the instances in which our judgments and decisions deviate from appropriate norms of rationality are explained by the fact that, in making these judgments and decisions, people rely on heuristics like representativeness “which sometimes yield reasonable judgments and sometimes lead to severe and systematic errors” (Kahneman and Tversky 1973 , 48). Though (1) and (2) have been challenged in a number of ways, they are both relatively modest reactions to the experimental findings. 3 However, many writers have suggested a much stronger and more disturbing conclusion, which maintains that people use these heuristics because they have no better tools available for dealing with many reasoning and decision making problems. According to Slovic, Fischhoff and Lichtenstein, for example, “It appears that people lack the correct programs for many important
judgmental tasks.We have not had the opportunity to evolve an intellect capable of dealing conceptually with uncertainty” (1976 , 174). What they appear to be suggesting is that: (3) The only cognitive tools that are available to untutored people are normatively problematic heuristics such as representativeness. This pessimistic conclusion seems to be endorsed in passages like the following in which Kahneman and Tversky, the founders of the heuristics and biases tradition, maintain that people use representativeness and other normatively defective heuristics not just in some or many cases but in all cases—including those cases in which they get the right answer: “In making predictions and judgments under uncertainty, people do not appear to follow the calculus of chance or the statistical theory of prediction. Instead, they rely on a limited number of heuristics which sometimes yield reasonable judgments and sometimes lead to severe and systematic errors” (Kahneman and Tversky, 1973 , 48). In light of passages like this, it is hardly surprising that both friends and foes of the heuristics and biases tradition suppose that it is committed to the claim that, as Gerd Gigerenzer has put it, “the untutored mind is running on shoddy software, that is, on programs that work only with a handful of heuristics” (1991b , 235). After describing the feminist bank teller experiment, the eminent biologist Stephen J. Gould, who is an admirer of work in the heuristics and biases tradition, asks: “Why do we consistently make this simple logical error?” His answer is: “Tversky and Kahneman argue, correctly I think, that our minds are not built (for whatever reason) to work by the rules of probability” (1992 , 469). 4 And, making the point a bit more bluntly, Lola Lopes, a psychologist who has been a trenchant critic of the heuristics and biases tradition, has suggested that researchers in that tradition think “that people are pretty stupid” (Lopes, quoted in Bower 1996 ).
3. Evolutionary Psychologists' Critique of the “Shoddy Software” Hypothesis The hypothesis that the only cognitive tools available to most human reasoners and decision makers are “shoddy software” like the representativeness heuristic has been the main target of an important critique of the heuristics and biases tradition mounted by evolutionary psychologists. In this section we'll give an overview of that critique. Though the interdisciplinary field of evolutionary psychology is too new to have developed any precise and widely agreed upon body of doctrine, there are three basic theses that are clearly central. The first is that the mind contains a large number of special purpose systems—often called “modules” or “mental organs.” These modules are invariably conceived of as a type of computational mechanism, namely, computational devices that are specialized or domain specific. Many evolutionary psychologists also urge that modules are both innate and present in all normal members of the species. While this characterization of modules raises lots of interesting issues—about which we have had a fair amount to say elsewhere (Samuels 1998 ; Samuels, Stich, and Tremoulet 1999 ; Samuels 2000 )—in this essay we propose to put them to one side. The second central thesis of evolutionary psychology is that, contrary to what has been argued by Fodor (1983 ) and others, the
end p.287 modular structure of the mind is not restricted to input systems (those responsible for perception and language processing) and output systems (those responsible for producing actions). According to evolutionary psychologists, modules also subserve many so-called central mental capacities such as reasoning, desire formation and decision making. 5 The third thesis is that mental modules are adaptations—they were, as Tooby and Cosmides have put it, “invented by natural selection during the species' evolutionary history to produce adaptive ends in the species' natural environment” (Tooby and Cosmides 1995 , xiii).
The Frequentist Hypothesis If much of central cognition is indeed subserved by cognitive modules that were designed to deal with the adaptive problems posed by the environment in which our primate forebears lived, then we should expect that the modules responsible for reasoning would do their best job when information is provided in a format similar to the format in which information was available in the ancestral environment. And, as Gigerenzer has argued, though there was a great deal of useful probabilistic information available in that environment, this information would have been represented “as frequencies of events, sequentially encoded as experienced—for example, 3 out of 20 as opposed to 15% or p = 0.15” (1994 , 142). Cosmides and Tooby make much the same point: “Our hominid ancestors were immersed in a rich flow of observable frequencies that could be used to improve decision-making, given procedures that could take advantage of them. So if we have adaptations for inductive reasoning, they should take frequency information as input” (1996 , 15–16). On the basis of such evolutionary considerations, Gigerenzer, Cosmides, and Tooby have proposed and defended a psychological hypothesis that they refer to as the Frequentist Hypothesis: “Some of our inductive reasoning mechanisms do embody aspects of a calculus of probability, but they are designed to take frequency information as input and produce frequencies as output” (Cosmides and Tooby 1996 , 3). This speculation led Cosmides and Tooby to pursue an intriguing series of experiments in which the “Harvard Medical School problem” used by Casscells et al. was systematically transformed into a problem in which both the input and the response required were formulated in terms of frequencies. Here is one example from their study in which frequency information is made particularly salient: 1 out of every 1000 Americans has disease X. A test has been developed to detect when a person has disease X. Every time the test is given to a person who has the disease, the test comes out positive. But sometimes the test also comes out positive when it is given to a person who is completely end p.288
healthy. Specifically, out of every 1000 people who are perfectly healthy, 50 of them test positive for the disease. Imagine that we have assembled a random sample of 1000 Americans. They were selected by lottery. Those who conducted the lottery had no information about the health status of any of these people. Given the information above: on average, How many people who test positive for the disease will actually have the disease? out of . (p. 24) In sharp contrast to Casscells et al.'s original experiment, in which only eighteen percent of subjects gave the correct Bayesian response, this problem elicited the correct Bayesian answer from 76 percent of Cosmides and Tooby's subjects. A series of further experiments systematically explored the differences between the problem used by Casscells et al. and the problems on which subjects perform well, in an effort to determine which factors had the largest effect. Although a number of different factors affect performance, two predominate: “Asking for the answer as a frequency produces the largest effect, followed closely by presenting the problem information as frequencies” (Cosmides and Tooby 1996 , 58). The most important conclusion that Cosmides and Tooby want to draw from these experiments is that “frequentist representations activate mechanisms that produce bayesian reasoning, and that this is what accounts for the very high level of bayesian performance elicited by the pure frequentist problems that we tested” (59). As further support for this conclusion, Cosmides and Tooby cite several striking results reported by other investigators. In one study, Fiedler 1988 , following up on some intriguing findings in Tversky and Kahneman 1983 , showed that the percentage of subjects who commit the conjunction fallacy can be radically reduced if the problem is cast in frequentist terms. In the “feminist bank teller” example, Fiedler contrasted the wording reported in section 1 with a problem that read as follows: Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice and also participated in anti-nuclear demonstrations. There are 100 people who fit the description above. How many of them are: bank tellers? bank tellers and active in the feminist movement? (p. 125) end p.289 In Fiedler's replication using the original formulation of the problem, 91 percent of subjects judged the feminist bank teller option to be more probable than the bank teller option. However, in the frequentist version, only 22 percent of subjects judged that there would be more feminist bank tellers than bank tellers. In yet another experiment, Hertwig and Gigerenzer (1994 ; reported in Gigerenzer 1994 ) told subjects that there were two hundred women fitting the “Linda” description, and asked them to estimate the number who were bank tellers, feminist bank tellers, and feminists. Only 13 percent committed the conjunction fallacy.
Studies on overconfidence have also been marshaled in support of the frequentist hypothesis. In one of these Gigerenzer, Hoffrage, and Kleinbölting (1991 ) reported that the sort of overconfidence described above can be made to “disappear” by having subjects answer questions formulated in terms of frequencies. Gigerenzer and his colleagues gave subjects lists of 50 questions similar to those described in section 1, except that in addition to being asked to rate their confidence after each response (which, in effect, asks them to judge the probability of that single event), subjects were, at the end, also asked a question about the frequency of correct responses: “How many of these 50 questions do you think you got right?” In two experiments, the average overconfidence was about 15 percent when single-event confidences were compared with actual relative frequencies of correct answers, replicating the sorts of findings we sketched in section 1. However, comparing the subjects' “estimated frequencies with actual frequencies of correct answers made “overconfidence” disappear.Estimated frequencies were practically identical with actual frequencies, with even a small tendency towards underestimation. The “cognitive illusion” was gone” (Gigerenzer 1991a , 89).
The Cheater Detection Hypothesis In section 1 we reproduced one version of Wason's selection task on which most subjects perform very poorly, and we noted that, while subjects do equally poorly on many other versions of the selection task, there are some versions on which performance improves dramatically. Figure 15.2 is an example from Griggs and Cox 1982 . From a logical point of view, this problem would appear to be quite similar to the problem in section 1, but the content of the problems clearly has a major effect on how well people perform. About 75 percent of college student subjects get the right answer on this version of the selection task, while only 25 percent get the right answer on the other version. Though there have been dozens of studies exploring this “content effect” in the selection task, the results have been and continue to be rather puzzling since there is no obvious property or set of properties shared by those versions of the task on which people perform well. end p.290
Figure 15.2 However, in several widely discussed papers, Cosmides and Tooby have argued that an evolutionary analysis enables us to see a surprising pattern in these otherwise bewildering results (Cosmides 1989 , Cosmides and Tooby, 1992 ). The starting point of their evolutionary analysis is the observation that in the environment in which our ancestors evolved (and in the modern world as well) it is often the case that unrelated individuals can engage in “non-zero-sum” exchanges, in which the benefits to the recipient (measured in terms of reproductive fitness) are significantly greater than the costs to the donor. In a hunter-gatherer society, for example, it will sometimes happen that one hunter has been lucky on a particular day and has an abundance of food, while another hunter has been unlucky and is near starvation. If the successful hunter gives some of his meat to the unsuccessful hunter rather than gorging on it himself, this may have a small negative effect on the donor's fitness, since the extra bit of body fat that he might add could prove useful in the future, but the benefit to the recipient will be much greater. Still, there is some cost to the donor; he would be slightly better off if he didn't help unrelated individuals. Despite this, it is clear that people sometimes do help non-kin, and there is evidence to suggest that nonhuman primates (and even vampire bats!) do so as well. On first blush, this sort of “altruism” seems to pose an evolutionary puzzle, since if a gene that made an organism less likely to help unrelated individuals appeared in a population, those with the gene would be slightly more fit, and thus the gene would gradually spread through the population. A solution to this puzzle was proposed by Robert Trivers (1971 ), who noted that while one-way altruism might be a bad idea from an evolutionary point of view, reciprocal altruism is quite a different matter. If a pair of hunters (be they humans or bats) can each count on the other to help when one has an abundance of food and the other has none, then they may both be better off in the long run. Thus organisms with a gene or a suite of genes that inclines them to engage in reciprocal exchanges with non-kin (or “social exchanges” as they are sometimes called) would be more fit than members of the same species without those genes. But of course, reciprocal exchange arrangements are
vulnerable to cheating. In the business of maximizing fitness, individuals will do best if they are regularly offered and accept help when they need it but never reciprocate when others need help. This suggests that if stable social exchange arrangements are to exist, the organisms involved must have cognitive mechanisms that enable them to detect cheaters and to avoid helping them in the future. And since humans apparently are capable of entering into stable social exchange relations, this evolutionary analysis led Cosmides and Tooby to hypothesize that we have one or more mental modules whose job it is to recognize reciprocal exchange arrangements and to detect cheaters who accept the benefits in such arrangements but do not pay the costs. In short, the evolutionary analysis leads Cosmides and Tooby to hypothesize the existence of one or more cheater detection modules. We call this the cheater detection hypothesis. If this is right, then we should be able to find some evidence for the existence of these modules in the thinking of contemporary humans. It is here that the selection task enters the picture. For according to Cosmides and Tooby, some versions of the selection task engage the mental modules that were designed to detect cheaters in social exchange situations. And since these mental modules can be expected to do their job efficiently and accurately, people do well on those versions of the selection task. Other versions of the task do not trigger the social exchange and cheater detection modules. Since we have no mental modules that were designed to deal with these problems, people find them much harder, and their performance is much worse. The bouncer-in-the-Boston-bar problem is an example of a selection task that triggers the cheater detection mechanism. The problem involving vowels and odd numbers presented in section 1 is an example of a selection task that does not trigger the cheater detection module. In support of their theory, Cosmides and Tooby assemble an impressive body of evidence. To begin, they note that the cheater detection hypothesis claims that social exchanges, or “social contracts,” will trigger good performance on selection tasks, and this enables us to see a clear pattern in the otherwise confusing experimental literature that had grown up before their hypothesis was formulated. When we began this research in 1983, the literature on the Wason selection task was full of reports of a wide variety of content effects, and there was no satisfying theory or empirical generalization that could account for these effects. When we categorized these content effects according to whether they conformed to social contracts, a striking pattern emerged. Robust and replicable content effects were found only for rules that related terms that are recogniza end p.292 ble as benefits and cost/requirements in the format of a standard social contract.No thematic rule that was not a social contract had ever produced a content effect that was both robust and replicable.All told, for non–social contract thematic problems, 3 experiments had produced a substantial content effect, 2 had produced a weak content effect, and 14 had produced no content effect at all. The few effects that were found did not replicate. In contrast, 16 out of 16 experiments that fit the criteria for standard social contractselicited substantial content effects. (Cosmides and Tooby, 1992 , 183) Since the formulation of the cheater detection hypothesis, a number of additional experiments have been designed to test the hypothesis and rule out alternatives. Among
the most persuasive of these are a series of experiments by Gigerenzer and Hug (1992 ). In one set of experiments, these authors set out to show that, contrary to an earlier proposal by Cosmides and Tooby, merely perceiving a rule as a social contract was not enough to engage the cognitive mechanism that leads to good performance in the selection task, and that cueing for the possibility of cheating was required. To do this they created two quite different context stories for social contract rules. One of the stories required subjects to attend to the possibility of cheating, while in the other story cheating was not relevant. Among the social contract rules they used was the following, which, they note, is widely known among hikers in the Alps: (i) If someone stays overnight in the cabin, then that person must bring along a bundle of wood from the valley. The first context story, which the investigators call the “cheating version,” explained that there is a cabin at high altitude in the Swiss Alps, which serves hikers as an overnight shelter. Since it is cold and firewood is not otherwise available at that altitude, the rule is that each hiker who stays overnight has to carry along his/her own share of wood. There are rumors that the rule is not always followed. The subjects were cued into the perspective of a guard who checks whether any one of four hikers has violated the rule. The four hikers were represented by four cards that read “stays overnight in the cabin,” “carried no wood,” “carried wood,” and “does not stay overnight in the cabin.” The other context story, the “no cheating version,” cued subjects into the perspective of a member of the German Alpine Association who visits the Swiss cabin and tries to discover how the local Swiss Alpine Club runs this cabin. He observes people bringing wood to the cabin, and a friend suggests the familiar overnight rule as an explanation. The context story also mentions an alternative explanation: rather than the hikers, the members of the Swiss Alpine Club, who do not stay overnight, might carry the wood. end p.293 The task of the subject was to check four persons (the same four cards) in order to find out whether anyone had violated the overnight rule suggested by the friend. (Gigerenzer and Hug 1992 , 142–43) The cheater detection hypothesis predicts that subjects will do better on the cheating version than on the no cheating version, and that prediction was confirmed. In the cheating version, 89 percent of the subjects got the right answer, while in the no cheating version, only 53 percent responded correctly. In another set of experiments, Gigerenzer and Hug showed that when social contract rules make cheating on both sides possible, cueing subjects into the perspective of one party or the other can have a dramatic effect on performance in selection task problems. One of the rules they used that allows the possibility of bilateral cheating was: (ii.) If an employee works on the weekend, then that person gets a day off during the week. Here again, two different context stories were constructed, one of which was designed to get subjects to take the perspective of the employee, while the other was designed to get subjects to take the perspective of the employer.
The employee version stated that working on the weekend is a benefit for the employer, because the firm can make use of its machines and be more flexible. Working on the weekend, on the other hand is a cost for the employee. The context story was about an employee who had never worked on the weekend before, but who is considering working on Saturdays from time to time, since having a day off during the week is a benefit that outweighs the costs of working on Saturday. There are rumors that the rule has been violated before. The subject's task was to check information about four colleagues to see whether the rule has been violated. The four cards read: “worked on the weekend,” “did not get a day off,” “did not work on the weekend,” “did get a day off.” In the employer version, the same rationale was given. The subject was cued into the perspective of the employer, who suspects that the rule has been violated before. The subjects' task was the same as in the other perspective [viz. to check information about four employees to see whether the rule has been violated]. (Gigerenzer and Hug 1992 , 154) In these experiments, about 75 percent of the subjects cued to the employee's perspective chose the first two cards (“worked on the weekend” and “did not get a day off”) while less than 5 percent chose the other two cards. The results for subjects cued to the employer's perspective were radically different. Over 60 percent of subjects selected the last two cards (“did not work on the weekend” and “did get a day off”) while less than 10 percent selected the first two. These experiments, along with a number of others reviewed in Cosmides and Tooby 1992 , end p.294 are all compatible with the hypothesis that we have one or more Darwinian modules designed to deal with social exchanges and detect cheaters. 6
4. What Do These Results Show? Three Optimistic Answers What do these results tell us about the rationality of ordinary people's reasoning and decision making? Evolutionary psychologists have urged that their findings support the truth of three increasingly optimistic claims. First, they maintain, the data suggest that: (4) There are many reasoning and decision-making problems on which people's intuitive judgments do not deviate from appropriate norms of rationality. Second, they have argued that: (5) Many of the instances in which our judgments and decisions accord with appropriate norms of rationality are to be explained by the fact that, in making these judgments, we rely on mental modules that were designed by natural selection to do a good job at nondemonstrative reasoning when provided with the sort of input that was common in the environment in which our hominid ancestors evolved. Finally, evolutionary psychologists have also on occasion issued exuberantly Panglossian proclamations suggesting that
(6) Most or all of our reasoning and decision making is subserved by normatively unproblematic “elegant machines” designed by natural selection, and thus any concerns about systematic irrationality are unfounded. This optimistic view is suggested in numerous places in the evolutionary psychology literature. For example, the paper in which Cosmides and Tooby reported their data on the Harvard Medical School problem appeared with the title “Are humans good intuitive statisticians after all? Rethinking some conclusions from the literature on judgment under uncertainty.” Five years earlier, while Cosmides and Tooby's research was still in progress, Gigerenzer reported some of their early end p.295 findings in a paper with the provocative title “How to make cognitive illusions disappear: Beyond `heuristics and biases.'” The clear suggestion, in both of these titles, is that the findings they report pose a head-on challenge to (1), the weakest of the pessimistic conclusions that have been drawn from research in the heuristics and biases tradition. Nor were these suggestions restricted to titles. In paper after paper, Gigerenzer has said things like: “We need not necessarily worry about human rationality” (1998 , 280); “More optimism is in order” (1991b , 245); “Keep distinct meanings of probability straight, and much can be done—cognitive illusions disappear” (ibid); and he has maintained that his view “supports intuition as basically rational” (1991b, 242). In light of comments like these, many observers have concluded that the view of the mind and of human rationality proposed by evolutionary psychologists is fundamentally at odds with the view offered by proponents of the heuristics and biases program. 7
5. Who's Right? The “Middle Way” and Dual Processing Theories Though this is not the place to defend our view in detail, we are inclined to think that the correct conclusions to draw about the rationality of ordinary folk from the large and growing body of experimental findings about reasoning and decision making are not nearly so optimistic as (6) nor nearly so pessimistic as (3). 8 To begin, note that (1), which claims that people's intuitive judgments on many reasoning and decision making problems deviate from appropriate norms of rationality, and (4), which claims that there are many reasoning and decision making tasks on which people's intuitive judgments do not deviate from appropriate norms of rationality, are entirely compatible. Moreover, we believe that the evidence reviewed in sections 1 and 3, along with many other studies that might have been discussed, make an overwhelmingly plausible case that both (1) and (4) are true. People do make serious and systematic errors on many reasoning tasks, but they also perform quite well on many others. The heuristics and biases tradition has focused on the former cases, while evolutionary psychologists have focused on the latter. The pessimistic (2) and the optimistic (5) are both explanatory hypotheses that make claims about the sorts of psychological mechanisms and processes that underlie these two sorts of cases. And here again, there is no inconsistency—both (2) and (5) might well be true. We think that each of these explanatory hypotheses may indeed turn out to be true,
though each has serious competitors. There isstill much to learn about the mechanisms and processes subserving reasoning, and until more is known we think it would be premature to either accept or reject (2) and (5). Though (1), (2), (4), and (5) might all be true, it can't be the case that both (3) and (6) are true, since the former insists that the only cognitive tools most people have available are normatively problematic heuristics that will lead to systematic errors, while the latter maintains that most reasoning is subserved by normatively unproblematic “elegant machines,” and thus that we need not worry about human rationality. But while (3) and (6) can't both be true, they can both be false, and we believe they are. There is nothing in the heuristics and biases literature that supports the claim that problematic heuristics are the only reasoning resources people can draw on, nor does this literature provide us with any reason to think that normatively unproblematic mechanisms like those posited by evolutionary psychologists do not exist. On the other side, evolutionary psychologists certainly have offered no reason to think that all reasoning is subserved by normatively unproblematic modules. Indeed, if it is granted, as we think it should be, that people typically do poorly on a large and important class of reasoning problems, then it is clear that (6) is indefensible. We believe that the “middle way” we've been urging between the pessimism suggested by the heuristics and biases tradition and the optimism proclaimed by evolutionary psychologists is compatible with and perhaps made more plausible by a family of dual processing theories about the mental mechanisms underlying reasoning and decision making that have gained increasing prominence in recent years. 9 Though these theories differ from one another in many details, they all propose that reasoning and decision making are subserved by two quite different sorts of system. One system is fast, holistic, automatic, largely unconscious, and requires relatively little cognitive capacity. The other is relatively slow, rule based, more readily controlled, and requires significantly more cognitive capacity. Stanovich (1999 ) speculates that the former system is largely innate, emerged relatively early in human evolution and, as evolutionary psychologists suggest, has been shaped by natural selection to do a good job on problems like those that would have been important to our hominid forebears. The other system, by contrast, evolved more recently and, “although its overall function was no doubt fitness enhancing, it is the primary maximizer of personal utility” (Stanovich 1999 , 150). This newer system is more heavily influenced by culture and formal education and is often more adept at dealing with many of the problems posed by a modern, technologically advanced, and highly bureaucratized society. Since the new system requires more cognitive capacity, is more influenced by culture and education, and does not get used automatically, Stanovich hypothesized that there might be significant individual differences in people's ability and inclination to use it. More specifically, he reasoned, people with higher cognitive capacity, as measured by instruments like the Scholastic Aptitude Test (SAT), and end p.297 with a cognitive style that emphasizes “epistemic self-regulation” should do better on tasks that the old system was not designed to handle. Stanovich agrees with the evolutionary psychologists that many of the tasks studied in the heuristics and biases
tradition fall into this category. In extensive studies of these tasks he has shown that, while the average performance on these tasks is indeed quite poor, there are some subjects who give the answer that the Standard Picture suggests is normatively correct on many of these problems, and these subjects typically have significantly higher SAT scores and score higher on tests designed to detect cognitive styles that include epistemic self-regulation. If Stanovich and other dual process theorists are on the right track, then the unbridled optimism sometimes suggested by evolutionary psychologists is unwarranted, since most untutored people do indeed lack the capacity to deal with a wide range of problems that are important in a technological society. But the glum pessimism often associated with the heuristics and biases tradition is not warranted either. Since the fast, automatic, and evolutionarily older system requires little cognitive capacity, everyone has the capacity to deal rationally with many reasoning and decision making problems that were important in the environment in which we evolved. Moreover, since the new, slow, rule-based system can be significantly affected by education, there is reason to hope that better educational strategies will improve people's performance on those problems that the old system was not designed to deal with. This hope is encouraged by the findings of Nisbett (1993 ) and his colleagues showing that, on many sorts of reasoning problems, a little education goes a long way.
NOTES Parts of this chapter are based on earlier work, especially Samuels, Stich, and Tremoulet 1999 ; Samuels, Stich, and Bishop 2002 ; Samuels, Stich, and Faucher 2002 ; and Samuels and Stich 2002 . We are grateful for the many valuable discussions, comments, and criticisms occasioned by this earlier work. Special thanks are due to Michael Bishop, Stephen Downes, Luc Faucher, Peter Carruthers, Richard Foley, Gerd Gigerenzer, Alvin Goldman, Daniel Kahneman, Ernie LePore, Ernest Sosa, and Patrice D. Tremoulet. 1. For useful surveys of this literature, see Nisbett and Ross 1980 ; Kahneman, Slovic, and Tversky 1982 ; Dawes 1988 ; Piatelli-Palmarini 1994 ; Sutherland 1994 ; and Baron 2001 . 2. Though the Standard Picture is widely accepted among philosophers, psychologists, and especially economists, it is of course not the only account of rationality that might be used to assess the quality of people's reasoning and decision making. (See Samuels, Stich, and Bishop 2002 and Samuels, Stich, and Faucher 2002 for discussions of some of the main alternatives.) Nor, for that matter, is the Standard Picture without its end p.298 problems. First, as Goldman (1986 , 82) and Harman (1986 , chap. 2) have both pointed out, it is far from clear in what sense, if any, a normative principle of reasoning can be derived from the rules of logic, probability theory, or decision theory. Nor is it clear which rules of these formal theories our judgments and reasoning mechanisms must accord with in order to count as rational. Indeed, serious disagreements still persist over
which versions of logic, decision theory, and probability theory the correct principles of rationality ought to be derived from. (See, for example, Gigerenzer 1991a ). 3. One of the most carefully developed of these challenges is Adler's (1984 ) argument aimed at showing that the results in Tversky and Kahneman's “feminist bank teller” experiment do not support the claim that subjects are committing a systematic reasoning error. Rather, Adler maintains, Gricean principles of conversational implicature explain why subjects tend to make the apparent error of ranking (h) (Linda is a bank teller and is active in the feminist movement) as more probable than (f) (Linda is a bank teller.). Another important family of challenges argues that when interpreting data from an experiment on reasoning, advocates of the heuristics and biases program typically assume that there is a single best way of applying the norms of the Standard Picture to the experimental task. But this is not always the case. Gigerenzer 2000 , for example, argues that there are usually several different and equally legitimate ways in which the principles of statistics and probability can be applied to a given problem and that these can yield different answers—or in some cases no answer at all. If this is correct, then obviously we cannot conclude that subjects are being irrational simply because they do not give the answer that the experimenters prefer. For an extended discussion of these challenges, see Samuels, Stich, and Faucher 2002 . 4. While Kahneman and Tversky's rhetoric, and Gould's, suggests that untutored people have nothing but normatively defective heuristics or “shoddy software” with which to tackle problems dealing with probability, Piattelli-Palmarini goes on to make the even more flamboyant claim that this shoddy software is more likely to get the wrong answer than the right one: “We areblind not only to the extremes of probability but also to intermediate probabilities—from which one might well adduce that we are blind about probabilities. I would like to suggest a simple, general, probabilistic law: Any probabilistic intuition by anyone not specifically tutored in probability calculus has a greater than 50 percent chance of being wrong” (Piattelli-Palmarini 1994 , 131–32). Despite passages like this, we think a case can be made that (3) is not really a central commitment of the heuristics and biases research tradition. For our defense of this claim, see Samuels, Stich, and Bishop 2002 . 5. The conjunction of the first two central theses of evolutionary psychology constitutes what might be called the Massive Modularity Hypothesis. For more on this hypothesis, see Samuels 1998 ; Samuels, Stich, and Tremoulet 1999 ; and Samuels 2000 . 6. Despite this evidence, the hypothesis remains very controversial. Many authors have proposed alternative hypotheses to explain the data, and in some cases they have supported these hypotheses with additional experimental evidence. Among the most prominent alternatives are the pragmatic reasoning schemas approach defended by Cheng, Holyoak, and their colleagues (Cheng and Holyoak 1985 and 1989 ; Cheng, Holyoak, Nisbett, and Oliver 1986 ) and Denise Cummins's proposal that we posses an innate, domain specific deontic reasoning module for drawing inferences about “permissions, obligations, prohibitions, promises, threats and warnings” (Cummins, 1996 , 166). Still other end p.299
hypotheses that purport to account for the content effects in selection tasks have been proposed by Oaksford and Chater (1994 ), Manktelow and Over (1995 ), Fodor (2000 ), and Sperber, Cara, and Girotto (1995 ). 7. In Samuels, Stich, and Bishop 2002 , we argue that, despite the rhetoric, (6) is not a central commitment of evolutionary psychologists who have studied reasoning. 8. For a much more systematic defense of the views offered in this section, see Samuels, Stich, and Bishop 2002 and Samuels, Stich, and Faucher 2002 9. Among those who advocate dual processing theories are Evans (1984 , 1989 ), Evans and Over (1996 and forthcoming ), Sloman (1996 ), Klein (1998 ), and Stanovich (1999 ). end p.300
chapter 16 GENDER AND RATIONALITY Karen Jones Reason and rationality have been the subject of feminist focus—much of it critical—right from the emergence of modern feminist thinking in the late seventeenth century (Astell [1701 ] 1970 , Masham 1705 , Wollstonecraft [1792 ] 1997 ). Nor is it surprising that feminist theorists should have turned their attention to reason and rationality, given that women's assumed deficiency in rationality has been used to justify their exclusion from full ethical standing and from full participation in education, politics, and the professions. Thus, addressing questions of gender and rationality came to seem central to the feminist liberatory project. In contemporary feminist theory, the claim is often made that reason and rationality are “gendered” or that reason is “male” or “masculine.” There might seem to be something inherently paradoxical in arguing—thus using reason and conforming to norms of rationality—that reason and rationality are male-biased and hence, at least from a feminist standpoint, suspect. As Haslanger writes: “I should also note that to my mind, there is something peculiar about engaging in discussion and reasoned debate over the value, or legitimacy, or reality, of reason and rationality. If there is something wrong with our commitments to reason, I doubt we'll find it this way (and I don't know what we could do about it if we did)” ([1993 ] 2002 , 211). Richards (1980 ) assumes that to critique rationality is to embrace a rhetorically dangerous and potentially self-undermining position: just as feminists cannot object to women's subordination while rejecting justice as a male-biased ideal, they cannot rebut sexist arguments while rejecting reason. Those in weaker positions should be especially wary of suggesting, as some feminists have (Greer 1970 , 108–9), that rational argument simply functions to serve the interests of the stronger. However, if, following Rawls's distinction between the concept of justice and conceptions of justice (1971, 5–6), we distinguish between the concept of rationality and various substantive conceptions of rationality, we can do much to dispel this air of paradox while at the same time providing a clearer way of stating the positions of those theorists, including Richards, who reject feminist critiques of reason. The term “rational” functions as a term of approbation with paradigmatic application to strategies for belief formation and action choice, and to the beliefs and actions so chosen.
Derivatively, the term applies to persons insofar as they conduct their practical and epistemic lives in accordance with the right norms for action choice and belief formation and insofar as they instantiate the character traits and virtues—whatever they are—that make them wise agents and responsible inquirers. Rational agents respond appropriately to their reasons, whether practical or theoretical. The concept of rationality is thus fundamentally normative. It is in virtue of their recognizing this normative role of rationality that theorists with very different understandings of what the right norms for action and belief are and what character traits and virtues are required in order to respond appropriately to one's reasons (i.e., theorists with very different conceptions of rationality) can nonetheless be said to be offering rival ways of explicating the same, fundamentally normative, notion. Feminist critiques of reason and rationality can be understood, nonparadoxically, as critiques of particular (often partial) conceptions of rationality—critiques typically, though not exclusively, undertaken on the grounds that that conception does not after all enable agents to respond appropriately to their reasons. Feminist stances toward gender and rationality divide into three broad camps: the “classical feminist” stance, according to which what needs to be challenged are not available norms and ideals of rationality, but rather the supposition that women are unable to meet them; the “different voice” stance, which challenges available conceptions of rationality as either incomplete or accorded an inflated importance; and the “strong critical” stance, which finds fault with the norms and ideals themselves, or with some subset of them. A variety of very different research projects, separated as much by what they presuppose about gender as by what they argue about rationality, are being pursued within this third critical camp and it will be the chief focus of this essay, but any adequate survey of the terrain requires that the other approaches be presented at least in outline. end p.302
I. The Classical Feminist Project Theorists pursuing this research project differ in their conceptions of rationality, but they are united in accepting the following three claims. 1.Available philosophical or commonsensical conceptions of rationality and of reason represent genuinely human norms and ideals. 2.The list of norms and ideals contained within available conceptions of rationality and of reason are sufficiently complete. They are sufficiently comprehensive to enable agents who follow them to respond appropriately to their reasons, whether practical or theoretical. That is, they fulfill the internal normative function of a conception of rationality, namely that of prescribing for wise agents and responsible inquirers. 3.The external normative functions assigned to rationality and reason are unproblematic. This claim requires further explanation. In addition to the (internal) normative role of making prescriptions for agents and inquirers, our capacity for rationality has been thought to ground our status as persons worthy of equal moral standing, and to be that
1.Available philosophical or commonsensical conceptions of rationality and of reason represent genuinely human norms and ideals. which separates us from (nonhuman) animals; moreover, the most worthwhile and distinctively human lives have been thought to consist in developing and exercising our rational capacities. Thus, our capacity for rationality has been accorded a normative significance not accorded to other human capacities and has been privileged ahead of those other capacities. The classical feminist project emerged first in the writings of Astell ([1701 ] 1970 ), Masham (1705 ), and Wollstonecraft ([1792 ] 1997 ) but continues to have advocates (Richards 1980 , Green 1995 , Pargetter 1986). Advocates of this approach respond to the charge that women are deficient in reasoning ability not by critiquing the yardstick against which women are held to fail to measure up, but rather by pointing out that some women do indeed surpass men in their abilities, and that shortcomings, where present, are the result of inadequate education. Thus Astell writes: “Can we Think and Argue Rationally about a Dress, an Intreague, an Estate? Why then not upon better Subjects? The way of Considering and Meditating justly is the same on all Occasions” ([1701 ] 1970 , 90, cited in Atherton [1993 ] 2002 , 32). And Wollstonecraft, somewhat tongue in cheek, compares women's reasoning ability with the equally deficient capacities of soldiers: “As proof that education gives this appearance of weakness to females, we may instance the example of military men, who are, like them, sent into the world before their minds have been stored with knowledge or fortified by principles.Soldiers, as well as women, practise the minor virtues with punctilious politeness” ([1792] 1997, 131). Astell, Masham, and Wollstonecraft each use their commitment to the central end p.303 importance of human reasoning capacities (point 3 above) to ground powerful arguments for women's education. Only by being educated can women develop their capacities for rationality and hence become full moral agents.
II. The Different Voice Project Proponents of this project, a project that has been largely pursued in connection with Gilligan's 1982 critique of Kohlberg's 1981 theory of moral development, deny claims 2 or 3 above (or both). That is, they claim that extant philosophical or commonsensical conceptions of reason and rationality are problematic not for what they include, but for what they leave out (McMillan 1982 , Ruddick 1980 ). Important capacities that enable us to respond well to our reasons are overlooked or devalued within traditional conceptions of rationality because those capacities have been associated with women and with women's nurturing activities within the private sphere, activities that have themselves been devalued. For example, some conceptions of rationality construe emotions as impediments to rational choice and to the reliable formation of true belief, but there is
evidence that our emotions enable us to respond to our reasons, both practical and theoretical (see Greenspan's contribution to this volume); thus the devaluation of those capacities traditionally viewed as feminine has resulted in a less than adequate conception of rationality (Jaggar 1996 ). It becomes a feminist goal to reinvest with value those capacities that have been associated with the feminine with a view to developing a more adequate conception of rationality that represents a genuinely human ideal.
III. The Strong Critical Project Projects in this camp, while otherwise heterogeneous in philosophical commitments, are united in denying the first claim constitutive of the classical feminist approach. That is, they deny that available philosophical or commonsensical conceptions of reason and rationality represent genuine human ideals. The ideals themselves contain some sort of male bias. Most often this bias is explicated in terms of those ideals being “gendered,” or “male,” or “masculine,” but, as we shall see, one might hold that the ideals or norms are biased without holding that end p.304 they are gendered. For example, they might be biased in the sense that they function to maintain relations of dominance and subordination between the genders by deauthorizing women as rational inquirers. Because the dispute among theorists in this camp is as much about gender as it is about rationality, it will be useful to classify theorists into rough family groups according to the conception of gender that underlies their claim that norms and ideals of rationality and reason are gendered. This will leave the position that the norms display some other sort of bias to be treated separately at the end of the section.
1. Gender as Symbolic In The Man of Reason (Lloyd [1984 ] 1993 ), her pathbreaking study of ideals of reason in the history of philosophy, Lloyd argues that “rationality has been conceived as transcendence of the feminine; and the `feminine' itself has been partly constituted by its occurrence in this structure” (104). That is, our understanding of both rationality and femininity are shaped by this exclusion, so that we understand rationality as requiring the transcendence of traits associated with femininity and we understand femininity in terms of that which has to be transcended or left behind in order to become “the man of reason.” It is important to be clear about the claim being advanced here, lest it be subject to ready refutation (see Grimshaw 1986 , 61–69). Lloyd is not claiming that there is, in the history of Western philosophy, a conception of reason or rationality, and that there is, in Western culture, a conception of femininity and that these are in opposition with each other, such
that the traits and virtues ascribed to the ideal inquirer are incompatible with feminine traits and virtues. This would be to make a mistake about both gender and the history of philosophy: just as there have been many different conceptions of the traits and virtues of the responsible inquirer, so there have been many different conceptions of feminine traits and virtues. Rather, Lloyd is claiming that whatever the conception of reason, the feminine and all it stands for has been excluded from it, and that this exclusion has shaped our understanding of both what it is to be an ideal inquirer and what it is to be feminine. Through readings attentive to metaphor as well as explicit argument, Lloyd identifies in the philosophical writings of thinkers as diverse as Augustine, Descartes, Hume, and Hegel, among others, a set of contrasts, including: reason/the senses, reason/the emotions, calm passions/regular passions, mind/body, culture/nature, knowledge/that which is known, universal/particular, theoretical/practical, and public/private. She argues that, either overtly in these texts, or covertly but revealed when a particular text is read in the context of the then-prevalent as end p.305 sumptions about gender, masculinity is aligned with the privileged and valued side of these contrasts and femininity with that which is inferior, or more subtly, complementary. Thus these contrasts are not value neutral: masculinity comes to signify all that is good, valued, and truly human, while femininity comes to be associated with that which is marginal, or not truly human, or apt to lead us astray. Moreover, these contrasts become available to symbolize the appropriate relationship between the sexes: just as reason should, in a well-ordered life, control the emotions, men should, in a well-ordered society, control women. Suppose we accept that historical conceptions of rationality have indeed been deeply inflected by gender metaphors, what follows from this? Lloyd does not claim that, just in virtue of being associated with traits that are thought of as masculine, our norms of rationality are thereby at fault qua norms of rationality. That is, her critique of these norms is not epistemic: it is a further question whether, for example, a conception of rationality that opposes reason to the emotions enables us to respond appropriately to our practical and theoretical reasons. But this might lead one to wonder how deep the critique is and why either philosophers or feminists should be concerned about it: perhaps “it mistakes for real features of reason what are in fact mere superficial accretions of metaphor in its philosophical articulation” (Lloyd [1984 ] 1993 , 74). Why should feminists be concerned with the various associations that have come to attach to the concept of “the feminine”—what, we might wonder, has any of this got to with actual women and with the feminist liberatory project? Even though there is no straightforward connection between the operation of “femininity” and “masculinity,” or “male” and “female,” as symbols in philosophical (or other) texts and the gender identities of actual men and women, the use of gender symbolism both is affected by and affects the social construction of gender identity (Lloyd [1984 ] 1993 , 74). The use of gender metaphors in a text can also be crucial to a text's argumentative structure by making some inferences seem compelling and others not; thus, metaphor
should not be thought of as “mere.” 1 Sometimes metaphors bridge what would otherwise appear to be gaps in argument (as perhaps associations between femininity and emotions have delayed recognition of the ways in which emotions contribute to rationality). In addition, the use of gender metaphors can make such texts available to serve the ideological function of maintaining current gender relations. This can happen even though the text does not itself contain any false assertions, such as false assertions about the capacities of women or about the capacities needed to respond well to reasons. As a parallel, consider Darwin's theory of evolution. Darwin appropriated metaphors of competition and struggle originating in Malthus's population theory and used them to articulate the theory of natural selection. These metaphors compromised neither the truth nor the well-groundedness of Darwin's theory, but they contributed to making that theory available for (mis)appropriation by social Darwinians and sociobiologists. 2 In these ways, gender metaphors can have flow-on effects both within a philosophical text and beyond it. Showing that a conception of rationality is gendered in the sense that it embeds gender metaphors in its articulation is not sufficient to show that it is inadequate as a conception of rationality—that is, that it makes a mistake about the norms, capacities and virtues needed to respond appropriately to reasons—though it rightly raises suspicions. Nevertheless, given the flow-on effects of such metaphors, they are of more than literary-critical interest and merit both philosophical and feminist scrutiny.
2. Gender as Psychosocial A number of different projects explore the connections between psycho-development and theory construction with a view to defending the claim that particular philosophical theories or particular cognitive stances toward the world are masculine. Drawing on object-relations theory, Bordo offers a reading of Cartesian objectivity as a defensive response to “anxiety over the separation from the organic female universe” (1987 , 5). Flax 1983 finds that themes in Western philosophy—including conceptions of rationality—are to be explained as the working through of tensions in male psychosexual development, tensions theorized in the work of object-relations theorist Chodorow (see Chodorow 1977 ). Keller argues that the perception that science is a distinctively masculine enterprise might be explained by science's commitment to a particular conception of objectivity, a conception that postulates a sharp subject-object split in which the knower is seen as radically separate from that which is known, and nature is objectified (1985 , 79). This conception of objectivity, with its emphasis on separation and autonomy, answers to the concerns of male psychosexual development. Both male and female infants must develop a sense of their own autonomy and agency through separating from the mother—a process fraught with anxiety for infants of either sex— but, in social contexts in which an infant's primary caregiver is a woman, the male infant must also differentiate himself from the mother in order to establish his gender identity. This gives rise to psychological differences between the genders: boys are required to individuate themselves from the mother more radically than are girls (Chodorow 1977 ).
These projects have been subjected to various critiques: Bordo's reading of Descartes is contested (Atherton [1993 ] 2002 ). The approach of using object-relations theory to ground generalizations about the masculinity of philosophy has been charged with inappropriately generalizing about gender: the conception end p.307 of gender that underlies this approach is alleged to be insufficiently historical and insufficiently attentive to the problem of trying to separate out the gender component from the complex identities of persons who are always also mem-bers of other socially marked categories (Grimshaw 1986 , chaps. 2 and 3; for a reply, see Bordo 1988 ). Even supposing that this account of gender does not generalize inappropriately, it focuses on gender difference rather than on gender domination (Bernick 1992 , 193). The critique of rationality to be examined next begins from an account of gender that focuses on relations of dominance and subordination.
3. Gender as Relational Catherine MacKinnon writes: The content of the feminist theory of knowledge begins with its criticism of the male point of view by criticizing the posture that has been taken as the stance of “the knower” in Western political thought.[That stance is] the neutral posture, which I will be calling objectivity—that is, the non-situated, distanced standpoint. I'm claiming that this is the male standpoint socially. I will argue that the relationship between objectivity as the stance from which the world is known and the world that is apprehended in this way is the relationship of objectification. Objectivity is the epistemological stance of which objectification is the social process, of which male-dominance is the acted-out social practice. That is, to look at the world objectively is to objectify it. The act of control, of which what I have described is the epistemological level, is itself eroticized under male supremacy. To say that women are sex objects is in this way redundant. Sexualized objectification is what defines women as sexual and as women under male supremacy. (MacKinnon 1987 , 50, cited in Haslanger [1993 ] 2002 , 227) In highly compressed form, this passage contains MacKinnon's account of gender, her account of the targeted conception of rationality (“objectivity,” “the non-situated, distanced standpoint”), and her account of the relationship between them. These elements of MacKinnon's theory and their interconnections are explored in Haslanger's “On Being Objective and Being Objectified” (Haslanger [1993 ] 2002 ), which offers a reconstruction of MacKinnon's argument for the claim that “objectivityis the male standpoint, socially.” First, the account of gender: according to MacKinnon, to be a woman is not to possess any particular set of psychological traits, nor is it to possess two X chromosomes or any other biologically defined trait or set of traits; rather, it is to have the relational and social property of being in a position of eroticized subordination: “Gender isa question of power, specifically of male supremacy
end p.308 and female subordination” (MacKinnon 1987 , 40); “Male and female are created through the erotization of dominance and submission. The man/woman difference and the dominance/submission dynamic define each other” (MacKinnon 1989 , 113). To be a woman is to be objectified—that is, to be seen and treated as an object for the satisfaction of (constructed) male sexual desire. To be a man is to occupy the social position of sexual objectifier. This social position of sexual objectifier is defined within power structures of dominance and subordination; thus being a sexual objectifier is not merely a matter of holding the objectifying attitude that women are for the satisfaction of male desire. Sexual objectifiers (men) have the power to make the sexually objectified (women) become who they see them as being (Haslanger [1993 ] 2002 , 225–26): “Men treat women as who they see women being”; “If woman is defined hierarchically so that the male idea of a woman defines womanhood, and if men have power, this idea becomes reality. It is therefore real. It is not just an illusion or a fantasy or a mistake. It becomes embodied because it is enforced (MacKinnon 1987 , 172, 119). Gender norms—norms of masculinity and femininity—identify standards of excellence for persons in the relevant gender role (Haslanger [1993 ] 2002 , 213–16). This account of gender norms contrasts with gender-as-symbol accounts of femininity, inasmuch as no amount of symbolic association between a trait and women is sufficient to make that trait count as feminine. It also contrasts with accounts of gender as psychosocial: we need not suppose that women have the psychological traits identified as feminine by a theory of psychosexual development. Some women may have these traits, some may not, but failing to have those traits is insufficient to escape the gender identity “woman” so long as one lives within social structures of eroticized subordination, and so may be treated as a woman (Haslanger [1993 ] 2002 , 246 n. 23). When gender roles are entrenched, the fundamentally prescriptive nature of gender norms is masked: people conform to those norms, and face sanction if they don't. Moreover, errant behavior is interpreted by those whose interpretations are taken to settle the matter through the lens of those norms, much as, for example, a slave's sabotage in tool-breaking is interpreted not as an expression of resistance, but as proof that slaves are “stupid” and suited only to menial labor. 3 According to MacKinnon, norms of objectivity have a role to play in this process: they lend support to the belief that women are by nature fitted to the socially subordinate position that they occupy. This belief, which projects onto women qualities that men desire that they have (e.g., docility and submissiveness), masks the fundamentally prescriptive nature of gender norms and thus lends stability to the oppressive relations constitutive of gender. Haslanger calls the relevant norm of objectivity “Assumed Objectivity” ([1993 ] 2002 , 233) and analyzes it as composed of a cluster of four subnorms linking end p.309 observations to claims about a thing's nature and claims about a thing's nature to prescriptions about how to treat it. The norms are as follows:
•epistemic neutrality: take a “genuine” regularity in the behavior of something to be a consequence of its nature. •practical neutrality: constrain your decision making (and so your action) to accommodate things' natures. •absolute aperspectivity: count observed regularities as “genuine” regularities just in case: (1) the observations occur under normal circ*mstances (for example, by normal observers), (2) the observations are not conditioned by the observer's social positions, and (3) the observer has not influenced the behavior of the items under discussion. (Haslanger [1993 ] 2002 , 232–33) But the norm of absolute aperspectivity would appear to be enough, if consistently followed, to alert an inquirer to the possibility that his own social position was affecting his observations. It would seem to be a way of weeding out projective beliefs, rather than lending support to them. To support the problematic projective belief, the norms require supplementation with: •assumed aperspectivity: if a regularity is observed, then assume that (1) the circ*mstances are normal, (2) the observations are not conditioned by the observers' social position, and (3) the observer has not influenced the behavior of the items under observation. (Haslanger [1993 ] 2002 , 233) Assumed aperspectivity in effect gives the inquirer epistemic carte blanche to ignore his social position and his influence on that which is observed. It, together with the three other norms of epistemic neutrality, practical neutrality, and absolute aperspectivity, jointly constitute the ideal of Assumed Objectivity. In the context of gender relations (i.e., of relations of eroticized domination and subordination), following this set of norms will justify the projective belief that women are, by nature, as men have made them become, which in turn justifies treating women as women, with all the violence and indignity that MacKinnon claims that involves. We thus arrive at the conclusion that following these norms of objectivity, in current social circ*mstances, is functional to the maintenance of gender hierarchy, and thus that the norms are, at the least, male-biased. Insofar as the projective belief is false, since it mistakes for women's nature something that is the result of social processes of force, the norms also lead us into epistemic error and so are not appropriate norms for responsible inquirers (Langton 1993 , 381–84). But we have not yet shown that the norms are gendered male; that is, we have not yet proven the claim that “objectivityis the male-standpoint, socially.” We must now investigate the connection between following those norms and functioning in the gender role of man—that is, in the role of sexual objectifier. end p.310
Norms can be more or less tightly connected to the social roles for which they are norms: they can be norms that are appropriate to the role, or norms the following of which is sufficient to function in the role in question. Haslanger ([1993 ] 2002 , 220) offers the example of norms for tenants: being considerate of neighbors is a norm for tenants, but clearly you can satisfy this norm without being a tenant. Paying rent on time, in contrast, is a norm for tenants the satisfaction of which is sufficient to make you a tenant. Similarly, there are two senses in which a norm might be said to be gendered: a norm might be appropriate to the given gender role, or satisfying the norm might be sufficient to function in that gender role. Call a norm “weakly gendered” if satisfying that norm helps one achieve success in the relevant gender role; call a norm “strongly gendered” if satisfying that norm is enough to function in that social role (222). Haslanger concludes that the norms of objectivity are weakly (contextually) gendered inasmuch as satisfying them in the context of relations of gender hierarchy contributes to success as an objectifier through justifying both objectifying beliefs and action, but they are not strongly (contextually) gendered. Women may follow these norms without thereby functioning as men, but if they follow them they collaborate in their own objectification (235). 4 It seems, then, that Haslanger has precisified the claim that rationality is gendered male in such a way as to enable a thorough evaluation of that claim. The evaluation of the claim proceeds on the basis of a specific conception of rationality, a conception given by the cluster of four norms that constitute Assumed Objectivity, and on the basis of a specific account of gender, an account that analyzes gender as a matter of social location in structures of eroticized dominance and subordination. The evaluation reveals a sense in which it is true that rationality is gendered male insofar as following the norms of Assumed Objectivity contributes to success as a sexual objectifier, but it does not support a strong claim identifying the standpoint of objectivity with the male standpoint. This precisification has, however, come at a price. First, as Haslanger herself notes ([1993 ] 2002 , 240), it is hard to see the resulting critique as a critique of traditional philosophical (or even commonsensical) conceptions of rationality. Once exposed, no one would endorse the cluster of norms together called Assumed Objectivity. Moreover, as Langton notes: “What is bad about Assumed Objectivity is not that it is objective, therefore male, therefore oppressive, contrary to MacKinnon. What is bad about the norm is that in a sense it is not objective enough” (Langton 1993 , 382). Indeed, the problem for the cluster of norms that constitute Assumed Objectivity emerges only when inquirers are given the epistemic license of assumed aperspectivity, and that norm, we might rightly think, is a norm that leads us away from, rather than toward, objectivity. We should attribute to MacKinnon a target conception of objectivity that lacks intuitive plausibility, lacks defenders, and identifies the problem as being one of insufficient objectivity when she identifies the problem as one of objectivity, only if a more plausible target for her critique cannot be found. Second, there are aspects of MacKinnon's critique of objectivity that are inadequately captured by a reconstruction that brings critical focus to the norm of assumed aperspectivity. In particular, MacKinnon seems to find fault with (something like) Haslanger's norm of absolute aperspectivity and not merely with the unwarranted epistemic license conferred by assumed aperspectivity. MacKinnon seems to think that
aperspectivity disparages the insights of feminist consciousness raising that reveal how one's social location can contributepositively to one's epistemic standing (MacKinnon 1989 , 83–105). In the next section, I explore an alternative reading of MacKinnon that focuses on the connections that MacKinnon draws between norms of objectivity and norms of credibility. This reading gives up on the project of arguing that norms of rationality are gendered male or are masculine; thus, it departs not only from Haslanger's reconstruction of MacKinnon, but also from symbolic and psychosocial accounts of how rationality might be gendered male. 5 The reading locates her critique of objectivity within a naturalist conception of how to defend norms and ideals that are norms and ideals for us—for the kinds of finite, embodied, socially located beings that we are.
4. Norms of Rationality and Norms of Credibility Agents and inquirers who conform to norms and ideals of rationality are those whose claims to knowledge must be taken seriously. The concept of “rationality” thus functions to separate agents into those fit to participate in ongoing conversation about what to do and what to believe, and those whose opinion can be overlooked or ignored. Many feminist critiques of rationality begin from observations of the ways in which norms of rationality function to deauthorize women as knowers. Women are dismissed as unreliable informants regarding their own experiences (Code 1995 , 58–82), as too emotional and too enmeshed in particular concerns to view practical problems objectively, as unable to follow abstract argument, and so on. Feminist contributions to public debate are likewise frequently dismissed as “unobjective” and “biased.” Norms of rationality thus ground norms of credibility. It is easy to underestimate the epistemic importance of norms of credibility if one has a conception of ideal knowledge seeking as an individual enterprise. On this conception of ideal knowledge seeking, responsible inquirers seek, as much as possible, to limit their cognitive dependence on others. Real knowledge is knowledge gained firsthand through experience or through independent consideration of argument (Locke end p.312 [1690 ] 1987 , I.iv.23). Reliance on testimony should be discharged, preferably through checking up on the content of that which is said, but, where this is impossible, through checking up on the credentials of the testifier. (And if this checking up requires further epistemic dependence, as it typically does when assessing expert credentials, then that dependence should in turn be discharged [Mackie 1969–70 , 254].) This individualistic conception of ideal knowledge seeking is inappropriate for the kind of cognitively and temporally limited beings that we are: our knowledge is made possible through epistemic dependencies on others. We routinely rely on others for information that we cannot acquire ourselves, in both everyday and formal contexts of inquiry. In science, the theory dependence of method (Boyd 1983 ) embeds epistemic dependency in at least the following ways: in the construction of scientific instruments (e.g., an imaging
technique might presuppose results from physics, chemistry, and computer science, results that it is practically impossible—and not good scientific practice—for the medical experimenter using the technique to check), in controlling for artifacts of the experimental design, and in the selection of what rival hypotheses to test against (where these are identified using sociological indicators such as whether the hypothesis is advanced by an appropriately credentialed research team, published in a reputable refereed journal, and so on). The acquisition of knowledge can progress as fast as it does only because such acquisition is a social rather than an individual enterprise; moreover, some things are knowable by finite creatures like us only because of cognitive specialization and teamwork (Antony 1995 , 72–73; Hardwig 1985 ; Hardwig 1991 ). The norms and social mechanisms that grant credibility to some and withhold it from others thus become proper objects of epistemological, and not merely sociological, study (Schmitt 1994 , Goldman 1999 ). Feminists contribute to this epistemological project by drawing attention to the ways in which assignments of credibility both reflect and reinforce gender hierarchy. MacKinnon writes: Having power means, among other things, that when someone says, “This is how it is,” it is taken to be that way. When this happens in law, such a person is accorded what is called credibility. When that person is believed over another speaker, what is said becomes proof. Speaking socially, the beliefs of the powerful become proof in part because the world actually arranges itself to affirm what the powerful want to see.Beneath this, though, the world is not entirely the way the powerful say it is or want to believe it is. If it appears to be, it is because power constructs the appearance of reality by silencing the voices of the powerless, by excluding them from access to authoritative discourse. Powerlessness means that when you say “This is how it is,” it is not taken as being that way. (MacKinnon 1987 , 164) 6 end p.313 Central to the feminist critique of rationality as functional to the maintenance of male dominance is the thought that extant norms of rationality—and in particular norms that determine who is and is not appropriately “objective”—underwrite discriminatory practices of credibility (Code 1991 , Code 1995 , Fricker 1998 , Jones 2002 ). The argument for this claim comes in two stages: first, it must be shown that norms of objectivity serve the ideological function of excluding women from “authoritative discourse”; second, the norms themselves must be found to be at fault, for if the fault lies merely in ideologically driven misapplications of the norms, then the solution would seem to lie in advocating a more rigorous adherence to, rather than rejection of, those norms. MacKinnon claims that the stance of objectivity consists of two ideals: distance and aperspectivity: “To perceive reality accurately, one must be distant from what one is looking at and view it from no place and at no time in particular, hence from all places and times at once” (MacKinnon 1989 , 97). To the extent that a putative knowledge claim can be shown to be the product of the inquirer's social situation, that claim is undercut as knowledge: “If social knowledge can be interpreted in terms of the social determinants of the knower, it is caused. Therefore its truth value, in this definition of the tests for truth,
is undercut. If it has a time or place—or gender—it becomes doubtful because situated” (MacKinnon 1989 , 98). MacKinnon claims that this requirement of aperspectivity is a “strategy of male hegemony” (MacKinnon 1982 , 537). The strategy has two faces, simultaneously legitimizing male claims to social knowledge and delegitimizing feminist claims to knowledge. Both faces of this strategy are illustrated in the debates surrounding research into innate differences between the sexes. I will focus on arguments from sociobiology that purport to show that gender hierarchy is an evolutionary consequence of the differential reproductive investments of human males and females, because the dynamic is especially close to the surface in these debates. 7 Trivers (1972 ) argues that the parental investment of females in their offspring is necessarily higher than the parental investment of males. Because the female must gestate the fetus, birth the child, and suckle it, females are limited in the number of offspring that they can produce. Evolution will thus favor female reproductive strategies that ensure the quality of the offspring, including strategies aimed at gaining access to resources for child rearing. Male reproductive strategies face different constraints: a male can impregnate numerous females. Males who mate frequently and with many different females will leave more—potentially significantly more—copies of their genes in subsequent generations. This puts males in competition for reproductive access to females, who thus constitute a limiting resource for males. Barash claims these evolutionary facts explain male aggressiveness, competition, and promis end p.314 cuity, as well as polygyny, and “female docility and sexual fussiness” (Barash 1979 , 89; see also Wilson 1978 . This basic adaptionist argument is bolstered with contested research on fetal androgenization and selected descriptions of animal behavior, behavior that has itself been described in terms of the gendered characteristics that it then lends support to. Thus, for example, Barash describes hermaphrodite worms as “active and aggressive when seeking to discharge their sperm, demure and discriminating when their more valuable eggs are at stake” (1979, 50, my emphasis). A scattering of anthropological observations complete the standard sociobiological package. Sociobiologists appeal to a fact/value distinction to deny a political agenda behind their research: “I worry that it will be misinterpreted and used as support for the continued oppression of women. My intent has been only to explore the evolutionary biology of male-female differences, not to espouse any particular social, political or ethical philosophy” (Barash 1979 , 89). Feminist objections to this research, in contrast, are seen as politically motivated and are explained away as the predictable consequence of feminist commitments. Feminists respond by pointing out the various ways in which socially available background beliefs shape sociobiological inferences: they lend support to the anthropological and animal studies that sociobiology calls on; they make the postulated psychological mechanisms look plausible even though many psychological mechanisms could have produced the hypothesized adaptive behavior under conditions of selection and rival psychological hypotheses are not ruled out. 8
The socially available background beliefs that lend plausibility to sociobiology are themselves shaped by gender relations; thus, these beliefs are to be explained by features of the theorists' social situation. It is simply not true that sociobiology conforms to the ideal of “the ostensibly non-involved stance, the view from a distance and from no particular perspective, apparently transparent to its reality” (MacKinnon 1982 , 538). However, given that the beliefs that feminists contest are relatively entrenched, it will tend to be feminists and not sociobiologists who are called on to defend the presuppositions of their position (Antony [1993 ] 2002 , 132–34). Thus credibility is differentially apportioned between feminist and mainstream views on gender. 9 We have completed the first stage of the argument for the conclusion that norms of rationality underwrite discriminatory credibility practices. That is, we have shown that in contexts of gender subordination (and race and class subordination—the point will generalize) norms of aperspectivity and distance will tend to be used ideologically. But the argument remains incomplete. We have not yet been given a reason to reject the norms of objectivity themselves, for the trouble seems to lie in their unfair application: feminists are held to standards that mainstream research does not meet. The problem lies in who is and who is not allowed to assume aperspectivity, rather than in aperspectivity itself. If this is the problem, end p.315 then perhaps the solution lies in holding mainstream research to the same standards as feminist research; indeed, the problem with research that reinforces gender relations seems to be that it is not objective enough (Antony [1993 ] 2002 , 208). The norm of aperspectivity could even be an important feminist tool, for all that has been said so far. Feminists can appeal to that norm to show that sexist science does not meet its own standards for objectivity. However, MacKinnon seems to think that the problem runs deeper; in particular, the norms of distance and aperspectivity are incompatible with the method of feminist consciousness raising. (In Haslanger's terms: the problem is with absolute aperspectivity (or something like it) and not with assumed aperspectivity.) Feminists cannot appeal to the norm of aperspectivity to engage in “ideology-busting” for their own epistemic practice—which they assume to be good epistemic practice—will not stand scrutiny by its lights. The norm of aperspectivity disparages knowledge claims that can be explained as the result of the inquirer's social location. Non-universal properties of an inquirer, such as social location, particular interests, and specific emotional responses are seen as distorting or “subjective” factors. Subjective factors should be controlled for and insulated from affecting the outcome of inquiry. Inquiry that does not adequately control for the influence of these factors is dismissed as “unobjective.” The ideals of neutrality and distance likewise recommend insulating inquiry from potentially distorting influences such as political commitments. Consciousness raising violated norms of aperspectivity, distance and neutrality. It was self-conscious in its commitment to political change. It was committed to “claim[ing] as valid the experience of women, the major content of which is the devalidation of women's experience” (MacKinnon 1989 , 116). While the formats of consciousness
raising groups—a grassroots phenomenon chiefly of the 1960s and 1970s—differed, they focused on recounting women's day-to-day experiences, especially of intimate relationships, and on their emotional responses to those experiences. They found this personal realm to be political insofar as it was the domain in which the power relations that constitute gender were constructed, maintained, and exercised. In women's often inchoate responses to their day-to-day experiences were found the resources with which to understand women's social position. These responses of, say, anger and discontent, responses that were by the lights of prevailing assumptions about gender roles unjustified, were not defended as veridical by following a procedure that gave equal weight (or indeed any weight) to the hypothesis that they were symptoms of a merely personal discontent (Scheman 1980 ). Rather, they were assumed to be justified responses to at least some aspect of women's social position. This assumption was important, since without the body of feminist theory that emerged from consciousness raising, from feminist political action, and from subsequent academic work on gender, feminists would not have been able to rebut the then-prevailing interpretations of women's experience. Given that this method starts out from a detailed examination of women's lived experience, an experience both available because of and constitutive of women's gender subordination, it finds social location to be an epistemic asset rather than a liability: “The process identifies the problem of women's subordination as a problem that can be accessed through women's consciousness, or lived knowing, of her situation. This implicitly posits that women's social being is in part constituted or at least can be known through women's lived-out view of themselves” (MacKinnon 1989 , 95–96). Thus, if we believe that feminist consciousness raising gave rise to knowledge about women's social position, then we must reject aperspectivity, neutrality, and distance as norms for inquirers. 10 Those norms are rejected not merely on account of the fact that they can be misused and so function to maintain gender relations. Even fairly applied, they deauthorize feminist claims to knowledge insofar as they judge the method of consciousness raising illegitimate as a method of rational inquiry. Moreover, following these norms would prevent women from acquiring social knowledge; thus following these norms hinders rather than helps inquirers respond to their reasons. In this way, the conception of rationality is rejected as inadequate qua conception of rationality. If consciousness raising enables women to learn the truth about gender relations not despite but because it starts from presuppositions, including most centrally the presupposition of the at least partial veridicality of women's emotional responses, then that must be because those presuppositions are true. Research agendas that begin from true presuppositions can further knowledge. This insight is not new to feminist theory: the theory dependence of method means that inquiry must always contain presuppositions. What matters is whether those presuppositions are true (Antony [1993 ] 2002 ). Bringing the issues of credibility and of consciousness raising to the center of a reading of MacKinnon's critique of norms of objectivity makes that critique more substantive than it is on Haslanger's reading. She is not attacking norms that, once made explicit, most people would repudiate. The norms of aperspectivity, neutrality, and distance, though somewhat vague, are part of a commonsense conception of rationality that continues to hold currency. If MacKinnon is right in thinking that methods like
consciousness raising can give rise to knowledge, then those norms must be rejected, for they cannot explain the success of that method. On this way of explaining the theoretical significance of feminist critiques of norms of rationality, they become part of the broader project of investigating what norms enable us—that is finite, social, epistemically dependent beings, with a specific cognitive architecture, functioning in particular environments—reliably to latch on to our reasons, both practical and theoretical. In focusing on the specifically political dimensions of credibility, especially as these involve gender, they contribute a distinctive perspective on the cluster of issues examined in social epistemology. In focusing on the epistemic importance of emotions, they join a end p.317 growing group of theorists who recognize the noncognitive grounds of some of our cognitive achievements.
NOTES 1. See Lloyd (1993 ) 2002 and Lloyd 1998 for a discussion of the significance of metaphor in philosophy. The role of gender metaphor in philosophical texts is being explored by a number of theorists working in the continental tradition; important contributions include: Irigaray 1985a and 1985b , and LeDoeuff 1991 . See also Alcoff 1995 and Whitford 1988 . In a related project, feminist science scholars have carried out detailed case studies of the impact of gender metaphors on scientific inquiry; see especially Bleier 1986 and Keller and Longino 1996 . 2. Darwin is aware of the potentially misleading effect of these metaphors; he writes: “I should premise that I use the term Struggle for Existence in a large and metaphorical sense, including dependence of one being on another” ([1859 ] 1968 , 116). This comparison between feminist work on metaphor and rationality and critiques of metaphor in Darwin, is made by Sturr (1998 ). 3. Marlene Gorris's film A Question of Silence illustrates this power of defining away acts of resistance. See also Hoagland 1988 , chap. 1. It is a powerful theme in MacKinnon's work, about which I'll say more in the next section. 4. It is worth noting that showing that a norm is weakly gendered is not enough to condemn it, even given the moral bankruptcy of gender roles. Kindness might contribute to one's excellence in the role of master. The role of master should be abolished, but it does not follow that kindness should be (Haslanger [1993 ] 2002 , 219). However, following the norm of assumed objectivity is incompatible with the feminist commitment to social change insofar as it legitimizes existing gender relations. 5. Thus, there will be passages in MacKinnon's writing that square poorly with this account. But that is true of Haslanger's reconstruction as well. What matters is whether the alternative reading brings out interesting features of MacKinnon's view that are obscured on current readings. I will argue that it does. 6. It is perhaps telling that Langton (1993 , 369), following a suggestion of Haslanger, substitutes “proven” for “proof” in the sentence, “Speaking socially, the beliefs of the
powerful become proof in part because the world actually arranges itself to affirm what the powerful want to see.” But that focuses on the truth making, rather than the silencing function, of power and thus elides the issue of credibility that is the central focus of this passage. 7. For an extended discussion of gender in sociobiology, with references to the larger literature, see Nelson 1990 . MacKinnon's own examples concern authority in social and legal contexts, including the ways in which such authority affects the definition of rape and makes the harms of p*rnography invisible. 8. For an overview of the issues here, see Gould 1981 , Kitcher 1985 . 9. Feminist critiques continue to be ignored in mainstream work on gender and end p.318 sexuality. For example, LeVay's much cited and praised book on sexual orientation (1993 ) uses contested work on androgenization and contested sociobiological arguments without so much as mentioning the feminist critique of that work. Norms and sociological mechanisms for determining credibility and thus determining who must be taken to be part of the ongoing discussion regarding gender and sexuality are here clearly functioning to feminist disadvantage. 10. There is a large feminist literature on the epistemic value of social location; see especially Hartsock 1998 , Code 1995 , and Harding 1991 . end p.319
chapter 17 RATIONALITY AND PERSONS Carol Rovane This chapter will explore eight related themes: (1) Persons are not merely rational, but possess full reflective rationality. (2) There is a single overarching normative requirement that rationality places on persons, which is to achieve overall rational unity within themselves. (3) Beings who possess full reflective rationality can enter into distinctively interpersonal relations, which involve efforts at rational influence from within the space of reasons. (4) A significant number of moral considerations speak in favor of defining the person as a reflective rational agent. (5) This definition of the person has led Locke and others to distinguish personal identity from animal identity. (6) Although it is a platitude that a person has special reason to be concerned for its own well-being, it is not obvious how best to account for that platitude. (7) Groups of human beings and, also, parts of human beings, might qualify as individual agents and, hence, as individual persons in their own rights. (8) There is a sense in which the normative requirements of rationality are not categorical but merely hypothetical. end p.320
1. What Kind of Rationality Do Persons Possess?
The most minimal condition that something must meet in order to qualify as rational is that it be an appropriate object of what Daniel Dennett (1978 ) calls the “intentional stance.” When we adopt the intentional stance toward something, we first attribute intentional attitudes to it and, then, we go on to predict the thing's behavior by working out what actions are rationally mandated by the attitudes we have attributed to it. Thus predictability via the intentional stance depends on conformity with the requirements of rationality. But it doesn't depend on understanding them. Indeed, if Dennett is right, it is appropriate to take the intentional stance toward things that couldn't possibly grasp the requirements of rationality, such as thermostats and pigeons. Other philosophers are not so liberal about what they are prepared to count as rational. Unlike Dennett, they insist that something cannot qualify as rational unless, in addition to conforming to the requirements of rationality, it also grasps those requirements and apprehends their normative force. In other words, a rational being must see that it ought to be rational. 1 Let us call this kind of rationality full reflective rationality. It is the kind that persons possess. Unlike thermostats and pigeons, persons have the idea that they ought to be rational. And this makes for a further difference. Unlike thermostats and pigeons, persons can qualify as rational even when they don't conform to the requirements of rationality— or, equivalently, even when their behavior is not predictable via Dennett's intentional stance. Being reflective, persons can inspect their own thoughts and actions and evaluate the extent to which they do and don't conform to the requirements of rationality. When they don't conform, they can respond to such rational failure by engaging in self-criticism and efforts at self-improvement. These self-critical activities show that persons are committed to being rational, and it is by virtue of this commitment that they can qualify as rational even in the face of rational failure. 2
2. What Are the Normative Requirements of Rationality? Although persons can be said to grasp the requirements of rationality and apprehend their normative force, they do not typically have explicit and articulate knowledge of what those requirements are. This should not be surprising, since it is a matter of significant philosophical and psychological controversy just what those requirements are. The knowledge that persons have of them is probably like the implicit linguistic knowledge that Chomsky (1966 ) claims human beings are born with, even though they can't (prior to linguistic theorizing) articulate it. But, however implicit their knowledge of the requirements of rationality might be, persons must nevertheless have some conception of them. Otherwise, they could not engage in the kinds of critical activities they clearly do engage in whenever they evaluate their own and others' thoughts and actions as being more or less rational. What follows is a very rough and schematic account of the sort of thing that persons must implicitly have in mind, by way of normative requirements of rationality, whenever they do engage in critical evaluation of their own and one another's rational performance. The most general normative requirement that rationality imposes on a person is the requirement to arrive at and act upon all-things-considered judgments. 3 These are
judgments about what it would be best for the person to do in the light of all of its beliefs, desires, and other attitudes. All-things-considered judgments are the outcome of a variety of rational activities that together comprise a person's deliberations, such as the following: resolving contradictions among one's beliefs, working out the implications of one's beliefs and other attitudes, ranking one's preferences in a transitive ordering. Each of these rational activities is directed at meeting a specific normative requirement of rationality. In the cases just mentioned, they are the requirements of consistency, closure, and transitivity of preferences, respectively. Deliberation involves many more rational activities, each of which is similarly directed at meeting some specific normative requirement of rationality. It is not important for the purposes of this chapter to identify all of these rational activities and the specific rational requirements they aim to meet. The important point is this: all of the rational activities that jointly comprise deliberation have a common purpose, which is to contribute to the overarching rational goal of arriving at and acting upon all-things-considered judgments. If this seems implausible, try to imagine what it would be like to arrive at all-things-considered judgments without satisfying the many other, more specific normative requirements of rationality. If a person refused to satisfy the requirement of consistency, for example, but persisted in believing a contradiction, then there might be no such thing as what it was best for the person to do in the light of all of its beliefs. One belief might direct the person to perform a certain action while its contrary directed the person not to perform it. Similar problems would arise if a person refused to satisfy the requirements of closure and transitivity of preferences. The person would be refusing to consider all things in the sense required for deliberation. Let us call the state that would be achieved if a person were to succeed in this endeavor of arriving at and acting upon all-things-considered judgments the state of overall rational unity. And let us note that there is one overarching normative requirement of rationality on persons that incorporates all end p.322 of the other, more specific requirements like consistency, closure, and so on, namely the requirement to achieve overall rational unity. It is important to bear in mind that all of these normative requirements of rationality— both the overarching requirement to achieve overall rational unity and the other, more specific requirements like consistency—apply only to individual persons and not to groups of them. This can be seen from our critical reactions. If one person believes two contradictory propositions, then we are bound to regard this as a failure of rationality on that person's part. But if one person holds one belief while another person believes its contrary, then we are not bound to regard either person as guilty of a rational failure. For each of them might have reasoned correctly from its own point of view, by arriving at allthings-considered judgments that take all of its background attitudes into account. Here is another way to put the point. Persons who disagree are not under the same kind of rational pressure to resolve their disagreements that individual persons are under to avoid inconsistencies among their attitudes. A person's commitment to being rational always requires it to resolve the latter, whereas a person's commitment to being rational does not always require it to resolve the former. In fact, it can happen that this commitment actually requires a person not to resolve the former. This happens when the person judges
that another person with whom it disagrees is definitely mistaken but, also, quite closed to changing their mind. To sum up: the normative requirements of rationality define what it is for an individual person to be rational. But this definition does not describe the actual rational performances of persons. Rather, it articulates an ideal of which persons can and typically do fall short. Persons are nevertheless committed to satisfying this ideal, at least insofar as they are committed to being rational. And this is so despite the fact that persons don't typically have explicit or articulate knowledge of it.
3. What Kind of Sociality Follows upon Reflective Rationality? Reflective rationality is essentially self-conscious. It may follow that it is a social achievement. Certainly, there have been many interesting attempts to show that this is so, including Hegel's master-slave dialectic, Mead's account of social self-consciousness, Wittgenstein's anti-private language argument, and Davidson's triangulation argument. 4 There is a common theme running through all of these end p.323 arguments, which is that we cannot achieve full self-consciousness unless we are led, through concrete social interactions, to regard ourselves from another's point of view. If any of the arguments were to succeed, it would provide us with a powerful antiskeptical result. We would be in a position to conclude that the sort of reflective being who can raise worries about whether it is in touch with an external world necessarily is in touch with it, because it necessarily is in touch with other minds that populate that world. Such antiskeptical arguments might appear to take us beyond the topic of this chapter, which is to explore the connection between persons and rationality. But the arguments are worth mentioning for the following reason. Many philosophers have taken for granted that we ought to define the person as a reflective rational being. If we were prepared to accept such a definition, and if, in addition, we accepted any of the above arguments as sound, we could conclude that persons are beings who are essentially social as well as rational. This is bound to have interesting ethical implications. Unfortunately, it is not possible to assess these difficult arguments here. But it is possible to pursue a related, if less ambitious, course. Rather than ask, Why are reflective rational beings necessarily social? we can ask instead, What kind of sociality do reflective rational beings possess insofar as they are social? In answering this latter question, we shall see that the kind of sociality in question is the kind that characterizes certain distinctively interpersonal relations. It should be noted first of all that there is a very abstract form of sociality that automatically follows upon reflective rationality. To be a reflective rational being is to have a conception of oneself as possessing a rational nature, and to conceive oneself in this way is to conceive oneself as belonging to a single kind that includes all and only those beings who share the same rational nature. This conception amounts to a form of sociality because it incorporates an understanding that all the members of the kind that it
picks out not only share a common nature but also know that they do—where this means that they know this of one another. Yet the way in which they know this of one another is so abstract that it barely qualifies as social knowledge. For it seems possible that they could have the knowledge without actually interacting with one another in the world. This, of course, is the very possibility that the arguments mentioned above are designed to rule out. They aim to show that the kind of self-conscious reflection that incorporates this abstract social knowledge that all reflective rational beings have of one another somehow presupposes that they also stand in concrete social relations in the world. Insofar as reflective rational beings ever do enter into concrete social relations with one another, they can bring this abstract social knowledge to bear in them. Their mutual knowledge of their common rational nature makes it possible for them to influence one another in rational ways, by offering reasons to one another. Whenever they do this, they leave it up to the other to consider the reasons end p.324 on offer from its own point of view; they leave the other free to reject those reasons or to accept them and, possibly, to act upon them. Like the description of rationality in the previous section, this description of rational influence is in certain respects idealized. Persons can be influenced in rational ways only insofar as they are implicitly committed to satisfying normative requirements of rationality—requirements that they inevitably fall short of satisfying and of which they have little explicit knowledge. However, precisely because the account is idealized in these respects, it is compatible with the existence of a whole lot of rational failure in the course of (attempted) rational influence among persons. It should not be assumed, therefore, that rational modes of influence are rare, or that they are confined to rarefied contexts like the academy or courts of law in which the various interlocutors are explicitly committed to providing reasoned arguments for one another's consideration. The sloppy and ill-considered interchanges of everyday life are also efforts at rational influence in the sense at issue. What distinguishes them as rational modes of influence is not that they involve perfect conformity to the requirements of rationality. What so distinguishes them is that they invite others to embrace certain attitudes or undertake certain actions by indicating that there is some reason to do so. These invitations do not presuppose that reflective rational agents are perfectly rational. Those who offer reasons and those who receive them can alike be prone to mistakes, laziness, confusion, etc. All that is required in rational influence is that those to whom reasons are being offered have the power, if they so choose, to take those reasons into account in their deliberations. It should be easy to see, therefore, that rational influence is ubiquitous in interpersonal relations. We might even go so far as to say that rational influence is in play whenever persons talk to one another. (There is one exception that ought to be registered: bullsh*t. But it seems fair to say that this exception is entirely parasitic on the normal case of genuinely rational influence.) 5 There is a crucial respect in which this account of rational influence is not at all idealized. It does not portray such influence as essentially respectful. The initial description provided above might make it appear that it is essentially respectful. For the description emphasized that when we offer reasons to others we leave it up to them to accept or reject
them. And this amounts to an acknowledgement that they are independent agents with their own points of view. However, this acknowledgement may fall well short of anything that most moral philosophers would want to call true respect for the freedom and autonomy of others. To see that this is so, consider the fact that lies and threats are species of rational influence. They are not like pushes and tugs that bring about their effects in purely physical ways that lie entirely outside the normative space of reasons. On the contrary. Liars offer their own—as it turns out unreliable—word as a reason for others to believe something—leaving it up to them to decide, from their own points of view, whether to believe the lies or not. Similarly, those who impose end p.325 threats on others also leave it up to them to decide, from their own points of view, whether to comply with the threats or not. Of course, a credible threat must be backed by a promise to make the victim suffer in some way if the threat is not complied with. But such a promise is hardly like a push or a tug. It too is an attempt at influence from within the normative space of reasons. Thus rational modes of influence are not always pure and good, nor are they always careful and cogent. What distinguishes them is that they are attempts to influence others from within the normative space of reasons, by exploiting and appealing to another's capacity for reflective rationality. 6
4. Is Reflective Rationality a Defining Condition of Personhood? The most powerful considerations that stand in favor of defining personhood in terms of reflective rationality are moral considerations. One moral advantage that follows upon defining persons in this way is it allows us to make the following equation: something is a person just in case it is the sort of thing that can be treated as a person. Here is why the equation follows. We noted above that rational modes of influence are ubiquitous in interpersonal relations. In fact, they are the distinguishing mark of all distinctively interpersonal relations. Whenever persons treat one another specifically as persons, they are engaging one another's points of view and interacting with them from within the space of reasons. We have just seen that this is so even when persons are disrespectful and abusive of one another. Even in such cases they are usually treating one another as persons as opposed to mere things, precisely because they are appealing to and exploiting their common rational nature. (Bullsh*t might appear to be an exception. Yet it seems that even in the course of bullsh*t persons are appealing to and exploiting their common rational nature, albeit without rational purpose.) There is no doubt that to embrace this definition and equation is to take a somewhat exclusionary view of persons. We would deliberately exclude from the kind “person” anything that we can't treat in distinctively interpersonal ways—for example, fetuses, the severely insane, the irretrievably comatose, and the hopelessly senile. But if this seems unduly exclusive, consider that the definition is highly inclusive as well. It entails that if
we can treat something as a person, then it is a person. And herein lies its moral advantage. For, whenever we find we can engage something in distinctively interpersonal ways—for example, in conversa tion and argument—then we cannot deny that we are confronted with a person. Any attempt at such a denial would be a form of prejudice. Moreover, it would be a hypocritical form of prejudice, because one would be explicitly denying someone's personhood while at the same time implicitly acknowledging it through interpersonal engagement. So, for example, it was necessarily hypocritical of the slave owners of the American South to deny that their slaves were persons, given that they implicitly acknowledged their personhood whenever they talked to them (and, even more revealingly, when they passed laws against their education). In contrast, it would not be hypocritical to deny that fetuses, the irretrievably comatose, the severely insane, and the hopelessly senile are persons (note that it would be nonsensical to pass laws against educating them). In making such a denial one would not be depriving these human beings of ethical significance. They would remain objects of affection, concern, respect, legal rights, and so on. All one would be doing is registering that there is an important ethical kind to which they do not belong, namely the kind that can engage one another in rational ways and, thereby, treat one another specifically as persons. So, one clear moral advantage of any view that makes reflective rationality a defining condition of personhood is that it helps to expose this kind of prejudice against persons, in which persons find it possible to treat certain others as persons and yet hypocritically deny that they are persons. It has obviously been of some real historical and political importance to expose and rectify this hypocritical form of prejudice against persons. 7 The temptation to succumb to hypocritical prejudice against persons is greater when weighty moral and political consequences follow upon acknowledging their personhood—as happens in any context where it is granted that rights and obligations attach to persons. But, of course, it is precisely when such rights and obligations are at issue that it is morally imperative not to succumb to the temptation. And there is no doubt that the sense of personhood that is typically at issue in philosophical defenses of such rights and obligations is the sense given by the definition in terms of reflective rational agency. This is true of Kant's moral theory and of all theories belonging to the social contract tradition that includes Hobbes, Locke, Rousseau, and, more recently, John Rawls. 8 Not only do these theories place persons at the center of our moral focus, but the considerations they offer for putting them at the center also have directly to do with reflective rational agency. It is qua reflective rational agents that persons are capable of contemplating Kant's categorical imperative and, also, of framing and evaluating social contracts with others like them. All of these theories provide us with moral considerations in favor of defining persons in this way. Locke ([1690b ] 1987 ) offered some other moral considerations that favor the definition, which speak to us somewhat independently of the various theories of rights and obligations. The moral dimension of personhood that was most central for him was accountability. He wanted to make sense of the idea that a person end p.327 could stand before God on Judgment Day and be held accountable for all its past actions. It is easy to see that this idea presupposes that persons are reflective rational agents. If the
Last Judgment is to be meaningful, the person judged must be able to apprehend its evaluative force. And this requires that the person have the very sort of capacity for selfcriticism that goes together with reflective rationality. In addition, Locke also held that the meaningfulness of the Last Judgment depends on another moral dimension of personhood besides accountability, namely prudential self-concern. He took for granted that the judgment must be backed by sanctions—the reward of eternal bliss or the punishment of eternal damnation. And these sanctions would be meaningless to the person being judged unless the person were prudentially concerned for its own wellbeing. Like accountability, this kind of prudential self-concern presupposes that persons are reflective rational agents. They must be reflective and rational enough to regard the threat of divine sanction as a reason for acting in one way rather than another. To sum up: we have reason to define personhood in terms of reflective rational agency if we agree with Locke that persons are, by nature, accountable for their actions and concerned for their own well-being; likewise if we embrace any of the standard theories of rights and obligations that attach to persons; and likewise if we think that there is something important about defining and exposing hypocritical prejudice against persons, in which we find it possible to treat something as a person and yet deny that it is a person.
5. Locke's Distinction between Personal and Animal Identity Locke ([1690b ] 1987 ) famously argued that the condition of personal identity is not the same as the condition of animal identity, and in doing so he inaugurated a philosophical dispute that has continued to the present. He took the distinction between personal identity and animal identity to follow from the definition of a person as a reflective rational being. However, before considering Locke's argument, it is important to see that, in principle, his “animalist” opponents can also accept that definition. Initially, it may seem that animalists cannot accept the definition. For, as we have already seen, the class of persons so defined does not coincide with the class of human beings (or, for that matter, any other suitably endowed animals). The reason why is that no human being possesses full reflective rationality at all stages end p.328 of its life, and some human beings do not possess it at any stage of their lives. But, according to animalists, it does not follow that a person is not a suitably endowed animal. On the contrary. As they see it, the capacity for reflective rational agency is part of the native biological endowment of human beings—though there is no reason why it couldn't belong to other animals as well. And they take it to follow that the identity of a person, qua thing with that capacity, is just the biological condition of identity for human beings and other suitably endowed animals. Thus, in their view, personhood is a status that suitably endowed animals achieve at certain times of their lives. But the thing that has the status—that is, the person—is always an animal whose life begins and ends with biological birth (or, perhaps, conception) and death. 9
In contrast, Locke held that the life of a person does not begin and end with these biological events. The life of a person is the life of a reflective rational being and, according to him, such a life begins and ends with the person's self-conscious awareness of it. Much of the rationale for Locke's view derives from his preoccupation with a proper understanding of how a person can stand before God on Judgment Day and be held accountable for all of its past actions. We noted above that it is a necessary condition for such accountability that a person be a reflective rational agent who can understand the normative force of God's judgment. Locke also insisted that a person must have first personal or self-conscious knowledge of its own past actions as its own. That is, the person must be able to regard all of its past actions as “mine.” Accordingly, when Locke considered the question of what God must call before Him when He calls all persons to be held to account, he reasoned that it was neither necessary nor sufficient that God should resurrect everyone's animal body. The only condition that need be satisfied is that each person stand before God with a consciousness of its past. Locke concluded that this must be the condition of personal identity, namely the condition in which there is a single, continuous consciousness. In order to drive the point home—that the life of a person, qua reflective rational agent who can be held accountable for its own actions, is not the life of an animal—Locke offered a thought experiment about a prince and a cobbler. He asked us to imagine that the consciousnesses of the prince and the cobbler are switched each into the other's body, and to consider who would be who after the switch. After the switch there would be two persons, each of whom could coherently be held accountable for its past actions. But we could not keep track of their identities in biological terms. If we did, we should have to say that the person with the princely body and the cobbling consciousness continued to be the prince and so ought to be held accountable for the prince's past actions, even though the person had no awareness of any princely past at all. Locke regarded this as absurd. He took it to be obvious that the person who could and should be held accountable for that princely past would be the person who was con end p.329 scious of that past as its own. And, in the case being imagined, this would be the person with the cobbling body into which the prince's consciousness had been switched. Contemporary neo-Lockeans have defended Locke's distinction by offering more thought experiments modeled on his about the prince and the cobbler. 10 One crucial innovation is that they have tended to exploit the other moral dimension of personhood that Locke took for granted—not personal accountability but, rather, prudential self-concern. They typically ask us to imagine that our own consciousness—where that includes all of our psychological attitudes and characteristics, such as our memories, beliefs, desires, commitments, and so on—is going to be switched into someone else's body and theirs into ours. Then they tell us that one of the resulting persons will be made to suffer while the other will be given all that it most wants and cherishes. And, finally, they ask us to choose, on the basis of prudential self-concern, which person will receive which fate. Anyone who chooses that the person who gets their original consciousness and psychological characteristics should be spared the harms and given the goods (and that the person with their original body should suffer the harms and be denied the goods)
thereby shows that they endorse Locke's distinction between personal and animal identity. Some have protested that Lockean-style thought experiments do not demonstrate anything more than a conceptual distinction between personal identity and animal identity. They maintain that it might turn out to be the case that, as a matter of scientific fact, a person's consciousness and psychological characteristics simply cannot be transferred from one human body (or brain) to another. 11 On this way of thinking, whether Locke was right to distinguish personal identity from animal identity is ultimately an empirical question. However, in the next two sections, we shall see that the idea of a person is, au fond, a normative notion. In a way, this has already begun to emerge. We saw in the previous section that the most powerful considerations in favor of defining personhood in terms of reflective rational agency are moral considerations. Locke himself was admirably clear on the point. Not only did he invoke moral concepts like accountability and self-concern in his arguments about personal identity, he also declared that the term “person” is a forensic term. Yet Locke certainly allowed that his normative starting point must eventually give way to a properly metaphysical investigation into the nature of persons qua things that have the normative significance that they do have. In our scientific age, it is natural to expect that such a metaphysical investigation might best be pursued by invoking the empirical methods of science. But we will see in the next two sections that there are ways in which persons, qua reflective rational agents, are not ordinary objects of metaphysical or scientific investigation, and that this is due to the way in which normative considerations are bound to impinge on any such investigation. end p.330
6. Some Puzzles about the Rational Basis of Self-Concern Prima facie, the recognition that some future state will be mine suffices to give me a special sort of reason to be concerned about it. We have seen that this normative assumption can hold of persons only insofar as they are reflective rational agents. We have also seen that the definition of persons as reflective rational agents does not suffice by itself—that is, without further supporting argument—to establish Locke's distinction between personal identity and animal identity. For animalists can agree that persons are reflective rational agents and yet deny that their lives can come apart from the life of a given animal in the way that Locke and neo-Lockeans have suggested. And, finally, we have noted that their dispute with Locke seems to be a purely metaphysical dispute that might be settled by future advances in science. However, the topic of prudential selfconcern lands us with a normative dispute as well, concerning what would constitute an adequate rational basis for a person's prudential self-concern. We shall see in this section that it is very difficult to insulate the metaphysical dispute about personal identity from this normative dispute. Both neo-Lockeans and animalists can agree that any adequate account of personal identity must show why it is that a person has reason to be prudentially concerned for its own well-being. But they would have to show this in quite different ways.
On their side, neo-Lockeans can argue that a person, qua rational agent, cannot fail to be concerned for its well-being for the following reason: the well-being of any rational agent will consist, in large part, in the satisfaction of its desires; and, so, since a rational agent is necessarily committed to acting so as to satisfy its desires it is, eo ipso, necessarily committed to acting so as to ensure its own well-being. As it stands, this account is not yet complete. For it has not explained why persons have reason to be concerned for their future well-being as well as their present well-being. After all, action always takes place in the present. And, so, it seems that the only desires that a person, qua rational agent, is necessarily committed to acting upon are its present desires. Yet neo-Lockeans can reasonably claim to find grounds for future-directed self-concern in their account of personal identity over time in terms of psychological continuity. Insofar as persons know that their lives will be characterized by such psychological continuity, they can frame and pursue long-term projects, including long-term relationships. In the context of such longterm projects and relationships, a person's present desires and activities aim to contribute to something that is temporally extended. And this means that the person's present concerns will be structured in a way that makes essential reference to its future concerns as well. 12 Obviously, animalists cannot claim to ground future-directed self-concern in quite the same way. If I conceive myself in animalist terms, then I must recognize that is possible for my life to go on without significant psychological continuity of the sort that would permit me to carry on with my long-term endeavors. From the neo-Lockean point of view, it would not be irrational for me to disregard a future to which I am not psychologically tied. And this is a crucial normative ground on which neo-Lockeans can argue that I ought not to regard such a future as mine. I should not regard a future to which I am not psychologically tied as mine precisely because, without such a psychological connection, there is no rational basis for the kind of self-concern that characterizes the life of a person. 13 However, animalists can also lay claim to a normative advantage over the neo-Lockean view, for they can insist that the neo-Lockean account of the rational basis of selfconcern faces a difficulty that they can avoid. The difficulty is that the neo-Lockean account fails to preserve the sui generis character of prudential self-concern. Here is why self-concern is not sui generis in the neo-Lockean account just sketched: it portrays a person's future-directed self-concern as grounded in its commitments to long-term projects and relationships. Such projects are often shared with others, and relationships always are. So although it is true that our commitments to them may give us reason to be concerned for our own future well-being, it is also true that such commitments will give us reason to be concerned for the well-being of others who also participate in them. That makes our future-directed self-concern analogous to our concern for those others and, hence, not sui generis. Animalists seem to have a somewhat better prospect of giving an account that preserves the sui generis character of self-concern. They can posit a natural instinct for self-love that is not grounded in any of a person's substantive commitments or in any of the psychological facts that neo-Lockeans take to constitute personal identity. In their view, it might simply be a biological fact that human beings evolved in such a way as to have a natural concern for their own well-being that is different in kind from their concern for all other things. 14
This is hardly the end of the matter. There are two ways in which neo-Lockeans can respond to the animalist argument just presented. First, they might question the rational pedigree of a biologically based instinct for self-love. Suppose that sustained critical reflection led us to see that the instinct was in tension with other values to which we are firmly committed. In that case, we might conclude that the instinct does not have an adequate rational basis after all, and that this undermines the alleged advantage of the animalist view. This first line of response can be backed by a second, which claims that there is actually an advantage in not portraying self-concern as sui generis, for to grant that there is an analogy between self-concern and concern for others constitutes a significant advance in our moral thinking. It is generally agreed that Sidgwick (1907 ) was the first to argue in this fashion, though many others have followed suit. Often, such arguments begin by pre end p.332 senting the analogy between self-concern and concern for others in the negative light cast by the “present aim theory of rationality.” As that theory's name suggests, all that rationality ever requires of us is that we deliberate from and act upon our present desires. One reason for embracing the theory is that it refuses to exaggerate the requirements of rationality. Another reason derives from reflection on the way in which deliberation and action always take place in the present. Since deliberation and action always do take place in the present, it seems that the only desires that it would be irrational not to take as the basis of one's deliberations and actions are one's present desires. It seems to follow from the present aim theory that we have no more reason to act in the present for our future well-being than we have to act for the well-being of others. It may have occurred to the reader that the neo-Lockean account of future-directed concern sketched above has already provided a solution to this problem for the present aim theory. This is a point to which we will soon return. But first, let's briefly consider some other responses that derive from more directly moral preoccupations. Sidgwick was not content to view the analogy between self-concern and other-concern in a negative light. He observed that almost everyone acknowledges the moral demands of prudence. And he urged that anyone who is prepared to extend their concern for their own well-being as prudence demands, beyond the present and into the future, ought also to be prepared to extend their concern beyond their own case to the case of others. Thomas Nagel (1970 ) takes a more Kantian perspective on the analogy, which leads him to seek a common ground of both prudence and altruism in the very structure of practical reason. According to him, if there is a rational ground of future-directed concern, it follows upon the existence of objective reasons that are not confined to our present desires because they are “timeless.” Similarly, if there is a rational ground of altruistic concern, it follows upon the existence of objective reasons that are not tied to our own points of view because they are “impersonal.” Nagel's Kantian strategy is controversial because it requires us to abandon the present aim theory of rationality. Derek Parfit (1984 ) pursues a strategy that is more similar to Sidgwick's. He presents the case of a Russian nobleman whose present desires revolve around certain radical political commitments. The nobleman anticipates that he will become conservative in old age, and he sees that if he were to take measures in the present to secure his future well-being,
they would require him to act against his present radical commitments. From both a phenomenological and an evaluative standpoint, there is a sense in which this nobleman finds his own future desires just as alien, and just as opposed to his own present wellbeing, as the desires of another might be. Although the present aim theory of rationality would license the nobleman to disregard his future desires, Parfit insists that he can do so only if he refuses to recognize any moral grounds for altruism. For, insofar as he acknowledges any moral pressure at all to be altruistically concerned for others, he must extend that concern to his own case. He is no more licensed to disregard end p.333 his own future desires than he is licensed to disregard the desires of others. Thus, while Sidgwick emphasizes that altruism cannot be any worse off than prudence, Parfit emphasizes in turn that prudence cannot be any be worse off than altruism. Earlier, we raised a worry about whether any account of self-concern that affords a close analogy with concern for others can be right, given that it fails to portray self-concern as sui generis. The philosophers we've just discussed are quite happy to find that selfconcern cannot be portrayed as sui generis, because they prefer a moral vision that demotes self-centered concerns in favor of the impersonal and impartial concerns of morality. The neo-Lockean account sketched above does not share that moral vision. It assumes nothing more than the present aim theory of rationality and goes on to argue that a person's present commitments can ground concern for its future well-being. This will be so insofar as a person is committed to long-term projects and relationships, for such commitments are bound to enlarge the scope of its self-concern beyond the present to include the future as well. In similar fashion, commitments to shared projects and relationships can enlarge the scope of one's concerns beyond one's own case to include others as well. But this kind of concern for others is quite unlike the sort of altruistic concern that Sidgwick, Nagel, and Parfit have in mind. It is neither impersonal nor impartial. It extends only to the particular persons with whom one is practically engaged through shared projects and relationships. When concern for others is partialist, it is much more like self-concern. So perhaps the worry that the neo-Lockean account fails to portray self-concern as sui generis is not so grave as it might, at first, have seemed. That all depends on just how close the analogy is, within a projects-based account of self-concern, between self-concern and concern for others. In the next section, we will consider the possibility that it is not an analogy at all but, really, the same thing. As we do so, we will continue to find ways in which metaphysical investigations into the person, qua reflective rational agent, cannot effectively be insulated from normative investigations.
7. Rational Unity as a Condition of Personal Identity It might seem paradoxical or, worse, nonsensical to say that the analogy between selfconcern and concern for others is not a mere analogy because they are really the same
thing. But there is a way of thinking about joint agency according to which it is neither paradoxical nor nonsensical. On this way of thinking, when end p.334 human beings exercise their agency together in joint endeavors, they can thereby constitute themselves as a single group agent. When this happens, what we would ordinarily think of as mutual concern on the part of the human constituents of the group agent might better be thought of as self-concern on the part of the group agent itself. The possibility of such group agents who are composed of many human beings goes together with another, which is the possibility of multiple agents within a single human being. By the lights of the definition of a person as a reflective rational agent, these amount to the possibilities of group and multiple persons. 15 This section will explore some reasons for granting that group and multiple persons truly are possible. We saw in section 2 that there is a single overarching normative requirement of rationality, which is the requirement to achieve overall rational unity by arriving at and acting upon all-things-considered judgments that take all of one's beliefs, desires, and other attitudes into account. The point at issue in this section concerns the boundaries within which overall rational unity is to be achieved. Since the requirement defines what it is for an individual person to be fully or ideally rational, these boundaries must, of course, be the boundaries that mark one person off from another. The animalist view of personal identity implies that rational unity ought to be achieved within the biological boundaries that mark one human being off from another. The neo-Lockean view implies that such unity ought to be achieved within the phenomenological boundaries that mark one consciousness off from one another. (In real life there is not much difference between these two views, since there tends to be a one-to-one correspondence between individual human beings and individual personal consciousnesses.) In this section, we are exploring some reasons for thinking that neither view is correct, because of the way in which human beings can exercise their native rational capacities in order to achieve different levels of rational unity within different boundaries. If a group of human beings managed to achieve overall rational unity at the level of the whole group, then the group itself would have satisfied the very normative requirement that defines what it is for an individual person to be rational. And, precisely because this is so, it could effectively be treated as an individual person in its own right. The same might occur within parts of human beings. Indeed, something very much like this occurs in certain dramatic cases of dissociative identity disorder, where several alter personalities emerge, each of which can be separately engaged in its own right. The claim that there could be such group and multiple persons is bound to meet with some skepticism. Unfortunately, it is not possible to give a full defense of it here. What follows is merely a rough indication of the kinds of considerations that support it. 16 Let us consider the group case first. It is well known that when human beings engage in group activities, their joint efforts can take on the characteristics of individual rationality. Think, for example, of marital partners who deliberate to end p.335
gether about how to manage their homes and families and other joint concerns. They may in the course of such joint deliberations do as a pair all of the things that individuals characteristically do in order to be rational: they may pool their information, resolve conflicts between them, rank their preferences together, and even arrive at all-thingsconsidered judgments together about what they should together think and do—where the “all” in question comprises all of their pooled deliberative considerations. The same can also happen in a less thoroughgoing way when colleagues coauthor papers, or when teams of scientists design and run experiments together, or when corporations set up and follow corporate plans. We tend to assume that such joint endeavors leave human beings intact as individual persons in their own rights. Insofar as that is so, it should be possible to engage those human beings separately in conversation, argument, and other distinctively interpersonal relations. But sometimes this is not possible. Sometimes marital partners won't speak for themselves. Their commitment to deliberating together is so thoroughgoing and so effective that everything they say and do reflects their joint deliberations and never reflects their separate points of view. The same can happen to coauthors, team members, and bureaucrats. The kind of case at issue here is not one in which human participants simply wish to give voice to the larger viewpoint of the groups to which they belong. Rather, it is a case in which the human constituents of the group are not committed to having separate viewpoints of their own. That is, these human beings are not committed to achieving overall rational unity separately within their individual lives. Yet it is not because they lack rational capacities. It is because those rational capacities are directed in a different way, so as to help fulfill a larger commitment on the part of a whole group to achieve overall rational unity within it. If this seems implausible, just think about two different attitudes you might bring to a department meeting. You might bring your own separate viewpoint to the table with the aim of convincing your colleagues to agree with you. This attitude takes for granted the status of each colleague as an individual person in its own right with its own separate point of view. The attitude also perpetuates that status, for the effect of adopting it will be that you maintain the separateness of your point of view by deliberating on your own, with the aim of achieving rational unity just within your own self. Even when you are moved by what your colleagues say, the reason why is not that you want to resolve disagreements with them, or do anything else that would help you to achieve rational unity as a group. You will be moved by your colleagues only insofar as what they say bears on your personal project of achieving such unity by yourself—by showing you that you have internal reasons, from your own point of view, to accept what they are saying. But you might bring a quite different attitude to a department meeting, one that would not perpetuate your separateness from your colleagues. You might bring to the table all you have thought of with respect to the issues the department faces, with a view to pooling your thoughts with your colleague's thoughts, so that you can together discover the allthings-considered significance of the whole group's thinking. If your colleagues do the same, then it won't be true that each of you is committed to achieving overall rational unity on your own; you will be jointly committed to achieving such unity within the department. And, for this reason, it will be possible for others to engage the department itself in conversation, argument, and other distinctively interpersonal relations. The department could be asked, for example, why did you do such and such, and there will be a coherent answer that reflects the department's joint deliberations. The point is not that
departments of philosophy typically have the commitment that would render them sufficiently unified to be engaged in this way. The point is only that it is possible. It is possible for human-size philosophy professors to undertake a commitment to achieve rational unity together. And, if they were to live up to that commitment, then the lines that divide one person off from another would have shifted. They would no longer follow the biological divisions that mark off different human animals, or the phenomenological divisions that mark off different centers of consciousness. They would follow nothing else than the commitment to rational unity that is characteristic of the individual person. To say that these lines can be redrawn in these ways is to say that the facts of personal identity are, quite literally, matters of choice. A similar argument can be produced for the possibility of multiple persons within a single human body. In fact, the argument is really a generalization of the argument for the possibility of group persons. What the argument shows, in effect, is that all cases of rational unity should be modeled on the unity of a group. Thus, the thought is that rational unity doesn't just happen as the inevitable product of some natural process, such as the natural biological development of a human being. Rational unity is something that is deliberately achieved for the sake of some further end. There are things that a philosophy department can do as a unified group person that no human-size person can do on its own. And that may constitute a reason why such human-size persons might initially decide to pool their efforts in a joint endeavor. If they implement their decision, they no longer maintain separate rational points of view. So, what perpetuates the group person once it has been brought into existence is not separate commitments on the part of its human constituents; it is up to the group itself to maintain its existence by continuing to strive for overall rational unity within it. When we view the unity of a human-size person along these lines, we must see it as deliberately achieved for the sake of some further end that couldn't be achieved without it. The appropriate contrast here is with an impulsive human being who doesn't strive for rational unity—who doesn't deliberate at all but simply follows current desires unreflectively and uncritically. Since the capacity to deliberate belongs to human nature, perhaps it is fair to say that such a human being is acting against its nature. But that doesn't harm the present point, which is that when human beings do exercise their rational capacities, they are generating rational unity end p.337 through their intentional efforts. And it is part of this same point that these capacities can be directed at the achievement of rational unity within different boundaries. An initially impulsive human being might come to strive for rational unity within each day, or week, or month, or year, or even a whole lifetime. The last goal was celebrated by Plato as part of the just life and by Aristotle as part of the virtuous life. In a less high-minded way, we now typically pursue the project of living a unified human life for the sake of other more specific projects such as lifelong personal relationships (friendships, marriages, families) and, also, careers. But what needs to be emphasized is that these are projects and they are optional. It is possible for human beings to strive for much less rational unity than these projects require and still be striving for rational unity. And, sometimes, the result may be relatively independent spheres of rational unity with a significant degree of segregation.
Such segregation is evident in degree in the lives of many human beings whom we find it possible to treat for the most part as roughly human-size persons. We may find, for example, that when we visit the corporation our friend “becomes” a bureaucrat who cannot recognize the demands of friendship at all. What this means is that our friend's life actually takes up a bit less than the whole human being we are faced with, the rest of which literally belongs to the life of the corporation. We often describe this phenomenon as a kind of “role playing.” But it can also be described as a fragmentation of the human being into relatively independent spheres of rational activity, so as to generate separate rational points of view that can be separately engaged. Of course, group endeavors do not necessarily result in such fragmentation; in principle, they can completely absorb the human lives that they involve (this may actually happen in the armed forces and in certain very intense marriages). But when a group endeavor does not completely absorb the human lives that it involves, there is a consequent split in those lives. And multiple persons can be conceived along just these lines. The only difference is that the separate rational points of view of multiple persons need not be imposed by involvements in group projects but, rather, by involvements in other sorts of projects that it is not possible for a single human being to pursue in a wholehearted and unified way. When a human being's projects are numerous, and when they have nothing to do with one another, this may make it pointless to strive to achieve overall rational unity within that human life. And it may be a rational response to let go of the commitment to achieving such overall rational unity within that human life and to strive instead for as many pockets of rational unity as are required for the pursuit of those relatively independent projects. So, just as a group person may dissolve itself for the sake of human-size projects that would otherwise have to be forsaken for the sake of the group's overall unity, a human-size person may dissolve itself for the sake of even smaller projects that would otherwise have to be forsaken for the sake of the human being's overall end p.338 unity. In such conditions, we will find the emergence of multiple persons within that human being, each of which can be treated as a person in its own right. The central idea of this section has been that the rational capacities of human beings can be exercised so as to achieve overall rational unity within different boundaries. If this idea is correct, then we ought not to think of the facts of personal identity as a metaphysical given. We ought to think of them instead as products of effort and will. As such products, persons would be things that exist for reasons in a sense that goes beyond merely having causes. They would be things that exist for the sake of further ends that their existence makes it possible to pursue.
8. Why the Requirements of Rationality Might Not Be Categorical The suggestions of the preceding section are surprising and, indeed, revisionist—not only with respect to the condition of personal identity but, also, with respect to the nature of the requirements that rationality places on persons. It is natural to think of those
normative requirements as being categorical in their demands. But the burden of the last section is that this is not so. Before explaining why this is not so, let us see why it at least seems so. Certainly, the command “Be rational” can speak only to rational beings. Conversely, rational beings are the sorts of things that, by nature, are responsive to that command. That is precisely what distinguishes them as rational beings to begin with, namely that they are responsive to the command. On the face of it, it is hard to see what other or further reason rational beings could possibly require in order to see the command as binding, than that they have a rational nature by virtue of which they can apprehend its normative force. Because this is so hard to see, it is hard to see how the requirements of rationality could possibly fail to be categorical in their demands. Even bracketing the arguments of the preceding section about group and multiple persons, there are some conditions in which it seems that the normative requirements of rationality might be set aside. However, we shall see that these are not best viewed as conditions on which the requirement to be rational is itself conditional. And, so, they don't serve to transform the apparently categorical imperative to be rational into a merely hypothetical imperative instead. end p.339 Consider, for example, Derek Parfit's (1984 ) discussion of the “irrationality pill.” He points out that taking such a pill might be the best way to avoid succumbing to a threat, since taking the pill would make us unresponsive to the rational force of the threat. Here is the hypothetical imperative on which we would be acting in such a case: If you can achieve the all-things-considered best outcome only by being irrational, then be irrational. This is irrationality in the service of rationality. And it hardly serves to bring out any interesting sense in which the end of being rational fails to be unconditional for rational beings. It would be far more significant if it could be shown that the commitment to being rational is conditional on having other more specific and substantive ends. It is certainly imaginable that a rational being might prefer to forgo engaging in rational activities. And this is imaginable even if the rational being regarded forgoing rational activity as a kind of suicide (because it would be the end of its life as a rational being). Now, if all this is imaginable, it is also imaginable that such a rational being might decide, upon reflection, not to commit this form of suicide because it had discovered that there are things worth doing—values worth pursuing and projects worth carrying out— that would require it to engage in rational activity after all. Such a discovery might be described as the discovery of a hypothetical imperative to be rational: if you find that there are substantive ends worth pursuing, then engage in whatever rational activities are necessary for the rational pursuit of them. A rational being to whom this hypothetical imperative speaks does seem to be one for whom the value of rationality itself is a substantive value that does not stand apart from other values and, so, is just as conditional in character as any other value might be. Yet there is still a sense in which the value of rationality can and must stand apart, even in such a case. The activity by which one reflects upon and comes to embrace the other substantive values in the light of which one can find reason to go on as a rational agent is already a rational process. And this means that a commitment to being rational is already implicitly presupposed in the very
evaluative process that is supposed to provide the conditional reasons for the commitment. It should be noted that there is no other value or end, besides the end of being rational, that necessarily figures this way in the life of a rational agent. These considerations do not amount to an ironclad argument to the effect that the normative requirements of rationality constitute a categorical imperative for rational beings. But they do strongly suggest that this is so. Why, then, do the arguments of the preceding section suggest otherwise, and how do they do so? The lesson is not that there is some deep mistake in the general direction of thought just outlined above. It is undeniable that the commitment to being rational is a necessary background condition for having other commitments to more substantive ends. It is also undeniable that a rational being is bound to be responsive to the normative requirements of rationality, on pain of losing its status end p.340 as a rational being altogether. But there is something that these thoughts fail to take into account, which is that in order to be rational, a rational being must achieve overall rational unity. Thus, the command “Be rational” is really equivalent to the command “Achieve overall rational unity.” And we have seen that the rational abilities of human beings can be exercised so as to achieve overall rational unity within quite different boundaries—within a whole single human being, or within a group of human beings, or within parts of human beings. And no matter what size persons already exist, it is within their power to redraw the boundaries between them by striving to achieve rational unity within different boundaries, either larger or smaller than the ones that currently individuate them. Once persons recognize that they face these alternatives, they need reasons to prefer one over another. Obviously, the command “Achieve rational unity” will not suffice to provide these reasons, because that command will be satisfied no matter which alternative is chosen. So persons must appeal to something apart from the bare requirements of rationality. They need to appeal to further, substantive ends, which would give them reason to achieve some given level of rational unity. As we put it at the close of the last section, persons exist for the sake of ends, which their existence, as the size agents they are, makes it possible to pursue. It follows that, when persons are commanded by rationality to achieve overall rational unity within themselves, we ought to allow that there is an implicit reference to and conditionalization on the ends for the sake of which they exist. The imperative, in other words, is an imperative to achieve overall rational unity within yourself, provided that you exist for the sake of ends worth pursuing and, furthermore, provided that you do not find other ends that are more worthy of pursuit, that would require you to strive to achieve rational unity within some larger or smaller boundary instead.
NOTES 1. Descartes expressed this view in The Discourse on the Method (in Descartes 1984a ). See also Davidson's “Rational Animals” in Davidson 2001 .
2. Defenses of such a “commitment”-based view of rationality (and, indeed, of mental attitudes generally) can be found in Bilgrami 2002 , Brandom 1994 , and Levi 1997 . It should be noted that such an idealized understanding of the normative requirements of rationality is somewhat controversial. On many accounts, there should be very little slippage between rational ideals and rational performance. This would include all accounts of the mind that portray mental attitudes as dispositions and all decision theories that take for granted “revealed preferences.” 3. This conception of rationality is so widespread that it would be impossible to list all of its prominent subscribers. It can be found in Aristotle's ethics and it is widely assumed in contemporary decision theory. 4. See Hegel 1977 , Mead 1913 , Wittgenstein 1953 , and Davidson 2001 's “Rational Animals.” 5. For a philosophical analysis of bullsh*t, see Frankfurt 1988 , chap. 10. 6. For a fuller account of the many forms that rational influence among persons can take, see Rovane 1998 , chap. 3.1. 7. I argued for this importance in Rovane 1998 , chap. 3.2. 8. See Kant 1785 , Hobbes 1968 , Rousseau 1968 , and Rawls 1971 . 9. For defenses of animalism, see Wiggins 1975 , McDowell 1999 , and Olson 1997 . 10. This literature is very large. Two prominent contributions are Parfit 1984 and Shoemaker 1984 . 11. This is the view expressed in Wilkes 1988 . 12. For an insightful discussion of the role of long-term planning in the deliberative lives of persons see Bratman 2000 . 13. This line of reasoning is rarely made explicit, but it is implicit in every neo-Lockean thought experiment about personal identity that invokes self-concern. 14. Something like this idea is manifest in virtually all of the British empiricist writings on morals—except, of course, that they make no reference to evolution. 15. Korsgaard 1989 raises these possibilities. 16. See Rovane 1998 for a more complete defense, as well as an explicit formulation of the analysis of personal identity that follows, which I call there the “normative” analysis of personal identity. end p.342
chapter 18 RATIONALITY, LANGUAGE, AND THE PRINCIPLE OF CHARITY Kirk. Ludwig Making sense of the utterances and behaviour of others, even their most aberrant behaviour, requires us to find a great deal of reason and truth in them. —Davidson 1984 , chap. 10, 153
1. Introduction
This chapter deals with the relations between language, thought, and rationality, and especially the role and status of assumptions about rationality in interpreting another's speech and assigning contents to her psychological attitudes—her beliefs, desires, intentions, and so on. Central to the discussion below will be the end p.343 status, in particular, of the Principle of Charity, first introduced by W. V. Quine as a maxim of translation, “assertions start[l]ingly false on the face of them are likely to turn on hidden differences of language” (Quine 1960a , 58–60). Donald Davidson has advocated a stronger form of the principle, which enjoins as necessary for interpretation of another's speech the assumption that she is largely rational and has largely true beliefs (Davidson 1984 , 27, 136–37, 152–53, 159, 168–67, 196–97, 200–201). The discussion will be organized around the following three questions: What is the relation between rationality and thought? What is the relation between rationality and language? What is the relation between thought and language? These questions are not independent. To possess a language is to be able to speak to another and to understand another's speech. One must therefore be an agent, something capable of acting, as opposed to merely moving or being moved, to possess a language. Language therefore presupposes thought and action. If rationality is a condition on thought and agency, as is widely (though not universally) assumed, then it is likewise a condition on possessing a language. In this case, seeing another as a potential interlocutor carries a commitment to finding her to be fundamentally a rational being, and to regarding oneself as likewise fundamentally a rational being. On the other hand, there is a long tradition that sees language as essential for rationality, but not for thought generally. Aristotle famously defined man as “the rational animal” (1984 : Topics, book 5, 132a22– 132a27; Nicomachean Ethics 1.7.1097b22–1098a20; De Anima, 1.5.645b14). We occupy an even more privileged position if, as Davidson has controversially argued (Davidson 1984 , chap. 11; 2001, chap. 7), language is a condition on thought, and rationality is essential to both: then thought, rationality, and language are possessed altogether or not at all. Section 2 takes up the relation of rationality to thought. Section 3 discusses the relation between rationality and the power of speech. Section 4 takes up the relation of thought to language. Section 5 summarizes the discussion.
2. Rationality and Thought
It is widely accepted that rationality is essential for thought. This section explains what this view comes to, the reasons for it in outline, and some of the objections that have been advanced against it. We begin with a brief characterization of what end p.344 is meant by “thought” in this discussion, and then of the domain and requirements of rationality. The term “thought” will be used to cover any psychological attitude with a propositional content. The term “propositional attitude” (coined by Bertrand Russell ([1922 ] 1961 ; [1918 ] 1985 ) will also be used interchangeably with “thought” in this sense. Central examples are beliefs and desires. A's belief that he is handsome has as its content that he is handsome; his desire to be admired has as its content that he is admired. Propositional attitudes are individuated by their psychological modes and contents. Thus, different attitudes can have the same content if they are entertained in different modes: one may have a belief that one will get well, for example, as well as a desire to get well. Other examples of propositional attitudes are intending, hoping, fearing, considering, wishing, and doubting. These are the psychological states especially relevant to a discussion of rationality because their contents, being propositional, can bear logical and semantic relations to one another; for example, one propositional content can require or support, or be incompatible or inconsistent with, the truth of another. From antiquity, the domain of rationality has been divided into the theoretical, having to do with the formation of belief, and the practical, having to do with the expression of agency (the terminology is due to Aristotle; for contemporary discussions see Audi 2001 and Harman 1999 , chap. 1). Theoretical rationality aims at arriving at true belief and avoiding false belief, nonhaphazardly. Practical rationality, to which theoretical rationality is an important aid, aims at getting what one most wants, in accordance with one's beliefs about what one can get and how one can get it (and, perhaps, though controversially, with evaluating one's ultimate ends; see, e.g., Audi 1990a and Brandt 1998 ). The degree to which someone is rational depends on the degree to which his attitudes exhibit patterns at and across times appropriate for ideal pursuit of his theoretical and practical goals. Theoretical rationality is concerned with having representational states that exhibit coherence at a time in the sense particularly of not displaying patterns that frustrate the goal of having true beliefs and avoiding false ones. Thus, for example, consistency in what one believes is an obvious goal of full or ideal rationality. Recognized inconsistency is worse than unrecognized inconsistency, though, and in cases in which it is difficult to discover the inconsistency we do not ordinarily count someone as irrational. (Frege's failure to recognize the inconsistency of his axioms for arithmetic does not convict him of irrationality.) Similarly for holding beliefs that, in the light of one's evidence, are not likely to be jointly true. Theoretical rationality concerns also how new beliefs are acquired in the light of new evidence, and with reasoning from, or acquiring new beliefs in the light of, beliefs which one already has. In general, the goal is to acquire true beliefs about, or relevant to, what one is interested in and to avoid false beliefs. Often what rationality requires is thrown into clearer relief by its break end p.345
downs. Thus, wishful thinking, believing something because you want it to be true, and arbitrary belief formation, believing for no good reason, are irrational, while apportioning belief to the degree of evidence is rational. (See chap. 1, this volume.) Practical rationality is also concerned with patterns of attitudes relevant to action, centrally belief, desire, and intention, both at and across times. An intransitive preference ranking is an example of an irrational pattern among conative states, since it can lead systematically to the frustration of one's practical interests (see chap. 10, this volume). If you prefer A to B, B to C, but C to A, then in principle you can be led to trade something of value (a penny, for example) in an endless cycle to get B for C, A for B, C for A and then again B for C, and so on, each pairwise trade seeming rational, though the entire set is not. Practical rationality concerns also the effective coordination of desires and beliefs in the pursuit of one's ends, which requires that one recognize what are the best means to ends most preferred and then implement them. Doing what one does not judge best all things considered—weakness of the will—is a familiar breakdown of diachronic practical rationality (see chap. 13, this volume, and Davidson 1980 , chap. 2). Similarly, though there is no general requirement on consistency in what one desires, there is a requirement on consistency in what one intends or plans to do, since inconsistent plans (the result of desires put through the sieve of practical reasoning) cannot be conjointly carried out (Bratman 1987 , chap. 8). Having the power of thought and action obviously does not require perfect rationality, whatever that could come to. Most of us are subject to all too familiar failings in both reasoning and acting. This gives point to seeing rationality as a normative requirement, as a standard by which to judge our thought and behavior. The question whether rationality is required for thought is whether something can be a thinking being without being largely rational, or, more generally, what the limits are on how irrational one can be and still be seen as capable of thought. Thus, the thesis that thought requires rationality can be put as the thesis that propositional attitudes can appear only in largely rational patterns, synchronic and diachronic. This is to say that the normative requirements of rationality, which tell us how we ought to reason, deliberate, and act, are also descriptive requirements on what it is to be a thinking being: we think and act largely as we ought, or we do not do so at all. The case for rationality being a requirement on thought rests on reflection on the conditions under which it is appropriate to attribute to something the basic attitudes of belief and desire, which are the primitive ingredients of agency. (We will assume that beliefs and desires come together or not at all—that is, that all thinkers are agents. While this might be challenged, there will not be space to discuss it adequately here.) Beliefs come only in appropriate patterns. To see something as having one belief requires seeing it as having many related beliefs (Davidson 1984 , chap. 14, 200; Davidson 2001 , chap. 7, 97–102; Stich 1983 , 53–60). We would not attribute to someone the belief that a gun was in the desk drawer except insofar as we see her as believing that guns are artifacts, fire bullets, and have barrels, that desks are solid, that drawers open, have space in them, and so on. These general beliefs are conditions on possessing the concepts that are involved in the particular belief in question and express basic relations that hold between those concepts and other concepts and conditions relevant to their application.
One must also typically have many beliefs about particulars to think a gun is in the desk drawer, which are supported by the general beliefs required to have the concepts, for example, that the desk is not alive, that it takes up space, that it is larger than the drawer, and so on. Action, which is the expression of agency, is seen in the light of both belief and desire. Agents do things. We wave to friends, we write letters, we prove theorems, build houses, cross the street. For present purposes, we can remain neutral on what actions are. Candidates are bodily movements, construed broadly to include certain mental events (Davidson 1980 , chap. 1), and causal processes (Dretske 1988 ). Actions are the products of intentions, which are formed in the light of our beliefs and desires. Typically the intention is formed on the basis of a desire for an end and a belief about how to achieve it. This shows the action in a favorable light, as done for a reason. Action explanations are often telescoped. We cite only the end or a connected means-end belief. We say, “He stepped on the brakes to stop the car,” or “He thought she'd be impressed by flowers.” The sense that an explanation has been given, however, depends on inferring that he thought stepping on the brakes would stop the car, or that he wanted to impress her. Davidson has argued influentially that every action is rationalizable by (made reasonable in the light of) a belief-desire pair that reflects means-end reasoning (Davidson 1980 , chap. 1). This has been disputed on the grounds that some actions are done for their own sake, so that no belief that it conduces to a further end is required to provide its reason (Locke 1974 , Mele 1988 , Mele 2003b ). In any case, each action can be represented as the correlate of the conclusion of a bit of practical reasoning about what it is best to do, and how to do it or what constitutes doing it. To see something as an agent then minimally requires seeing it as exhibiting coherent patterns of belief and a certain kind of reasonableness in acting. While these observations show that some minimal level of coherence in thought, desire, and action is necessary for something to be an agent, they don't by themselves guarantee that an agent cannot have many inconsistent beliefs, reason mostly ineptly, or act mostly on reasons that are not best all things considered. Further support for the view that agents' attitudes as a whole must be seen as appearing in a largely rational pattern lies rather in reflection more generally on the conditions under which we are willing to treat something as an agent. We should expect some unclarity about where to leave off calling something end p.347 an agent. Like most natural language terms, “agent” is semantically vague. Our practice does not determine a precise cutoff point along dimensions of variation relevant to its application. However, the degree to which a system can be seen as rational is clearly a relevant dimension of variation for the term “agent.” The less coherence we find in the set of attitudes we are thinking of as potentially those of an agent, the less clear we are that what we are considering is a possible agent at all. Moreover, reflection on cases—for example, step-by-step increases in overall incoherence—shows that to see another as an agent at all requires finding a large degree of coherence in his outlook, interests, intentions, and behavior. This coherence is expressed in seeing how the attitudes attributed both make for a reasonable picture of things, from the agent's point of view, and make sense of her behavior as an expression of agency. We do not treat anything as
an agent unless we find a large degree of reason in what it thinks and does. This can escape our notice because we tend to focus on follies that we can identify only in the light of the vast background of reasonable action and belief that makes sense of people having the attitudes we take them to have. Irrationality, seen in this light, is a “perturbation of reason,” not its absence (Davidson 2001 , chap. 7, 99). We identify an irrational action or attitude as one that is a departure from an otherwise largely rational pattern, a pattern that makes sense of the attribution of attitudes that depart from it. While this is the majority view, it has been challenged, nominally, at least, on both empirical and conceptual grounds. Investigation shows that these challenges, in addition to being subject to internal criticisms, involve a mischaracterization of the thesis. Thus some psychologists have argued that experimental results show that human beings are not in fact rational animals (Johnson-Laird and Wason 1977 , Nisbett and Ross 1980 , Tversky and Kahneman 1983 ). “Pace Aristotle,” one author says, “it can be argued that irrational behavior is the norm not the exception” (Sutherland 1994 , vii). Here is an example. Experiments show that most college students do poorly on some versions of the selection task. Consider four cards, with “E,” “C,” “5,” and “4” on their faces. Each card has a letter on one side and a number on the other. The task is to turn over the minimum number of cards to check whether, if a vowel is on one side, an odd number is on the other. Many subjects turn over the “E” card and the “5” card, but not the “4” card. The conditional is falsified if a card with a vowel on one side has an even number on the other, so the “E” and “4” cards should be turned over. These and other experiments have been alleged to show that most people reason irrationally on many simple reasoning tasks. However, it is clear that these observations about mistakes on reasoning tasks (as understood by the experimenter) do not undermine the view that it is constitutive of the propositional attitudes that they occur in largely rational patterns. Identifying what subjects believe and want to do in the experimental situations end p.348 itself requires seeing their attitudes appearing in characteristically rational patterns. Their failures to reason as well as they could are themselves identifiable only because we already see them as largely reasonable creatures, responding largely reasonably to the tasks set before them. At most, then, these experiments could show there is some standard of rationality most people fail to meet. But, as Ernest Sosa (1999 ) has noted, to infer from this that most people are irrational or not rational is like inferring from the fact that human eyesight falls short of some extrahuman standard that most of us can't see well or can't see at all. Apart from this, the interpretations of these sorts of experiments have been challenged on the grounds that they often involve overly simple assumptions about how the experimental subjects understand their task, and fail to distinguish between what subjects can do (competence) and what they do in the circ*mstances (performance). For example, we often use sentences to convey more than is conveyed by their literal meaning. We understand a card shark who says, “If it has this mark on the back, it's an ace,” to be implying also that all aces have this mark on the back. Otherwise the remark would not be to the point. This would explain why in the selection task subjects turn over the “5”
card. Inattention and nonsalience can help explain why many don't turn over the “4” card. Conditionals are typically used in modus ponens reasoning rather than modus tollens reasoning. Inattentive subjects then are apt to check inference potential first. However, they understand the mistake when it is explained to them. Thus, one must also distinguish between what one is capable of on reasoning tasks in principle and one's actual performance, which may be affected by a wide range of conditions. (See Cohen 1981 and Davidson 1980 , chap. 14, 270–73, for further discussion.) In addition to psychologists' efforts to show we are not in fact as rational as we might suppose, some philosophers have argued that there are only “minimal” conceptual constraints on the irrationality of agents. In particular, under this heading, it has been argued that there is no set of inferences or inference forms that all thinking beings must endorse, no “`fixed bridgehead' of true and rational beliefs” (Hollis 1982 , 73). The view that there is a fixed bridgehead of true and rational beliefs has been attacked in particular by Stephen Stich (1990 , chap. 2), who follows Cherniak 1986 in rejecting it on the grounds that we can imagine a people whose feasibility ordering for inferences is inverted with respect to ours: inferences we find easy, they find hard, and vice versa. Again, however, prima facie this is not a challenge to the thesis that propositional attitudes appear only in largely rational patterns. Nothing in the thought experiment suggests that the imagined people have massively inconsistent or incoherent beliefs or suffer from significant breakdowns of practical reasoning. In addition, the thought experiment that drives the argument is of doubtful coherence. The main difficulty is that the hypothetical others must possess the same concepts we do, for the complex inferences they are to find easy and we hard are end p.349 ones couched in terms of our concepts. But we do not attribute concepts to people who are not able to recognize the simplest of inferences they conceptually underwrite involving them. So anyone who possesses those concepts must be able to recognize the validity of the same simple inferences we do (Biro and Ludwig 1994 ).
3. Language and Rationality Given that agency and thought are necessary for language, if rationality is a condition on agency and thought, it is a condition on possessing a language. This section discusses how rationality is related to interpreting others as speakers. The next section takes up the question of whether language is necessary for thought. Rationality, Interpretation, and the Principle of Charity Speakers are agents. Hence, to possess a language is to be at least as rational as agents must be in general. A constraint then on interpreting another as a speaker, on assigning meanings to his sentences, and contents to his attitudes, is that the pattern of assignments makes him largely rational. Moreover, the attitudes attributed must make sense of him as an agent capable of performing speech acts and communicating with others on a potentially limitless range of topics. Because seeing another as rational is a matter of
finding his attitudes appropriately related by content and mode, this imposes a holistic constraint on interpretation, in the following sense. In finding patterns in another's behavior appropriate for interpretation as linguistic and other intentional behavior, one must be sensitive to the full pattern of assignments of meanings to sentences and attitudes in judging the appropriateness of the interpretation of any given bit of behavior. Radical interpretation is interpretation of another without the usual contingent assumptions of commonality of language, culture, or psychology (Davidson 1984 , chap. 9). Reflection on radical interpretation is a tool in investigating the most general and abstract requirements on having a language. When we strip away all the common aids to interpretation and consider how we could come to interpret another simply on the basis of whatever we can know a priori about speakers, and on behavioral evidence open to public observation, we uncover what patterns in behavior we must perforce think are there if it is to be interpretable as linguistic behavior. This helps to show what content our linguistic and allied concepts end p.350 have—those of meaning, truth, and the propositional attitudes—by showing how they are related to independent evidence for their application. The requirement that we see others as largely rational in order to interpret them has most famously been discussed as a constraint on radical interpretation under the heading of “The Principle of Charity.” We will concentrate here on Davidson's version of the principle. The principle has two distinguishable aspects, which are motivated differently. In his later work, Davidson has distinguished the two strands as the Principle of Coherence and the Principle of Correspondence (Davidson 1985a , 92; Davidson 2001 , chap. 14, 211). The Principle of Coherence is concerned specifically with the a priori requirements on seeing something as an agent that can perform speech acts. It enjoins one to find another to be largely epistemically and practically rational, on pain simply of not being able to see the other as an agent at all. The grounding for this part of the Principle of Charity is a priori reflection on the nature of agency and the propositional attitudes of the sort sketched briefly in the previous section of this chapter. Davidson does not attempt to spell out in general the requirements on seeing others as largely rational, but argues that in practice it is a matter of seeing another as rational and reasonable by one's own lights, adjusting for differences in position and interests. This is consonant with seeing the principle as appealing to a priori constraints. For to say something is rational by our lights is just to say that we see it as required of a rational agent given our concept, that is, the concept, of rationality. Thinking about rationality from the point of view of interpreting another (or seeing another as an agent), in the light of the other's behavior, sheds some additional light on the concepts of agency and of the propositional attitudes, for it helps to highlight what gives them their practical content, namely, the way in which they enable us to make systematic sense of the movements of an object, by seeing it as goal directed in the light of largely true beliefs about the environment. To the extent to which we cannot make good sense of a system in these terms, we should not see it as being an agent with propositional attitudes: the supposition that it has propositional attitudes in this case is idle, and disconnected from the role the concepts of the attitudes play in our
making sense of things. (See Davidson's work in the philosophy of action, in particular the essays collected in Davidson 1980 , esp. chaps. 1–3, 5, 12–14; also Davidson 1982 and Davidson 2001 , chap. 7.) The Principle of Correspondence has received three different formulations. To explain these and the motivation for the principle, it will be necessary to say more about Davidson's account of radical interpretation, for the Principle of Correspondence is introduced to solve a problem that faces the radical interpreter when his task is cast in a certain way. In particular, Davidson gives a central place in the interpreter's procedure to confirming for a speaker's language a formal recursive truth theory similar to the sort characterized by Tarski (1932) 1983 . The truth theory is to serve as the vehicle for a compositional meaning theory for the language (Ludwig 2002 ). The theory has as theorems sentences of the form (T) (or notational variants) (T) s is true-in-L iff p where “s” is replaced by a description of a sentence of the speaker's language in terms of its primitive significant parts, and “p” by a sentence in the interpreter's language. The radical interpreter aims to construct and confirm a truth theory that issues in theorems of this form by observing the speaker in his environment. Davidson has suggested that a truth theory confirmed from this standpoint can be used to interpret the speaker's utterances, that is, that we can use the sentence “p” used in the interpreter's language to give truth conditions for the speaker's sentence s in (T) to interpret s. This is in effect to hold that if a truth theory is confirmed from the radical interpreter's standpoint, “is truein-L iff” in (T) can be replaced with “means that.” (For natural languages, which contain context sensitive expressions, such as demonstratives and tensed verbs, the truth and meaning predicates must be relativized to features of context relevant to their interpretation.) The radical interpreter must determine the appropriate (T)-sentences for the speaker's language. Davidson supposes the radical interpreter can, from behavioral evidence, figure out what sentences of his language a speaker thinks are true (holds true). The radical interpreter can then identify the conditions in the environment in which the speaker holds true a sentence. For some sentences, whether a speaker holds them true or not will not be sensitive to what goes on in his environment (e.g., “2+2=4”). In other cases, there will be systematic correlations between the speaker's holding true a sentence and what's happening in his environment (as in the case of “That's a rabbit”). These latter are the key to what the speaker means by his words and what he thinks. A speaker's holding true a sentence is a “vector of two forces”: what he believes and what he means by the sentence (Davidson 1984 , chap. 10, 196). Specifically, Davidson assumes (idealizing somewhat) that if a speaker believes that p, and believes that a sentence s of his expresses that p (at the time), then the speaker infers that s is true. For the speaker is committed to its being true that p and is presumed to know that if it is true that p, and s expresses that p, then s is true. Accepting this, if we knew which sentences a speaker held true and what they meant, we could infer what he believed. Likewise, if we knew what the beliefs were that formed the basis for his holding true various sentences, we could infer what the sentences meant. We start out, however, knowing neither what he believes nor what his sentences mean. The role of the Principle of Correspondence is to show how to break into this circle. end p.352
The first of the three formulations the Principle of Correspondence has received is that other speakers are largely in agreement with one, explicable error and ignorance aside. The second is that when a speaker holds true a sentence, by and large the sentence is true. The third is that a speaker's beliefs, particularly those that are responses to his environment, are largely true. Each has some textual support. However, the third is the correct formulation, for that is the only one that points to a way of fixing either what someone believes or the meanings of his sentences. For if we can assume that a speaker's beliefs about his environment are mostly true, then we know that when his beliefs are correlated with things in the environment, his beliefs are likely to be about things they are correlated with. If we further assume that there are enough constraints overall on interpretation to narrow down to a single salient condition, for each belief, what it can be reasonably correlated with for interpretation, we can identify what he believes on the basis of the correlated conditions that prompt his beliefs, for that is what they are about. If the third interpretation of the principle is correct, the correctness of the first two follows, but not vice versa. The Principle of Correspondence is justified as a necessary condition on being able to construct a justified interpretation of another speaker on the basis of (i) a priori constraints imposed by our conceiving of him as an agent possessing a language, and (ii) information about his interaction with his environment. It plays a role similar to the Principle of the Uniformity of Nature, which holds that nature evolves in accordance with general laws, in Hume's account of what is necessary to justify our inductive inferences. We infer from past regularities to future regularities. For this to be reasonable, we must assume minimally that nature evolves according to general laws, which are reflected in past observed regularities, which can then be projected into the future. Suppose that we knew that we could be justified in believing things about the future on the basis of the past. Then we would be able to infer that it is reasonable to accept the Principle of the Uniformity of Nature, since if it were not, we could not be justified in our inductive inferences. The Principle of Correspondence is justified in a manner similar to this, by appeal to the assumption that we can come to a justified interpretation of another speaker on the basis of the evidence and a priori constraints available in radical interpretation, and the claim that if we were not justified in believing the Principle of Correspondence, we could not come to a justified interpretation of another speaker on the basis of the available evidence, given the constraints. The argument to justify the Principle of the Uniformity of Nature by appeal to our being justified in making inductive inferences is question begging. For we cannot know our inductive practices are reliable a priori, and any a posteriori justification of inductive practices must rely on the Principle of the Uniformity of Nature. Hume argued that the Principle of the Uniformity of Nature, since it end p.353 is itself a generalization that covers the past and future, and is not knowable a priori, could be justified only by appeal to induction, thus leaving both the principle and our inductive practices unjustified.
For the Principle of Correspondence to fare better than the Principle of the Uniformity of Nature, we must be able to know a priori that we can arrive at correct interpretations of any other speaker on the basis of radical interpretation. If we could know it only a posteriori, and the assumption of the Principle of Correspondence were necessary in any inference to what a speaker means by his words and what he believes, we could not support the Principle of Correspondence in this way without presupposing it. Davidson adopts the strategy of justifying the Principle of Correspondence by appealing to an a priori justification of the possibility of correctly interpreting any other speaker from the standpoint of the radical interpreter. The argument for the possibility of correctly interpreting any other speaker from the standpoint of the radical interpreter rests on the observation that language is by its nature a medium for communication, so that interpretation must be something that can be accomplished on the basis of evidence that is available interpersonally: “That meanings are decipherable is not a matter of luck; public availability is a constitutive aspect of language” (Davidson 1990 , 314). The central idea here is that it is of the nature of a language that it is a device that enables its speakers to communicate with others, and that this can be done only on the basis of interpersonal evidence for what others mean. Yet granting this does not seem to be enough to yield the result that Davidson needs. It would seem enough to satisfy the requirement that language be a medium for communication that it guarantee that if we had mostly true beliefs and were confronted with someone who speaks a language with largely overlapping expressive powers, and who believed similar things about the environment, we would be able to interpret him. Justifying the Principle of Correspondence requires the stronger claim that to have a language is to be interpretable correctly on the basis of public evidence in any circ*mstances by any possible speaker (see Lepore and Ludwig 2004, Ludwig 1992 , and Ludwig 1999 for further discussion). Even if we despair of justifying the Principle of Correspondence, we need not despair of being able to interpret others correctly. For we need not suppose that our epistemic resources are restricted to those available to the radical interpreter. We can appeal to knowledge of features of our own psychological type and to the fact that in practice others whom we want to interpret are conspecifics embodied in the same way we are and in similar environments, to infer with some plausibility the sorts of things they are apt to be thinking, in order to constrain our interpretations. This would not, however, yield any argument to the conclusion that the empirical beliefs of linguistic beings were by their nature largely correct. What is the relation between the two parts of the Principle of Charity? Traditionally, being rational has not been seen as requiring largely true empirical end p.354 beliefs. Thus, it has been supposed that one could be rational but mostly mistaken in one's empirical beliefs, someone, for example, who reasons perfectly, but who is systematically deceived about his environment by Descartes's Evil Demon, and, so, through no fault of his own. Rationality on this view is a precondition on having largely true empirical beliefs but does not require it. Rationality would require a large number of true general beliefs as a condition on possessing the concepts involved in any of an agent's beliefs, but
these would not be empirical. For example, to possess the concept of red, one would have to believe, indeed, to know, that red is a color, that red is a feature of the surface of an object, that no surface, viewed from one position, can be two different colors at the same time, that surfaces are extended, that extended objects occupy space, and so on. These propositions are not empirical propositions; they are necessary, and knowable a priori. If Davidson is right, one cannot be rational and a speaker without having mostly true empirical beliefs as well. To be a rational speaker, then, would be to be largely right about the world. Yet, while being largely right in one's empirical beliefs, if Davidson were right, would be necessary for being a rational speaker, it would not thereby be an aspect of being rational. I have characterized the Principle of Coherence as being grounded in a priori reflection on the nature of agents. However, thinking about its role specifically in the light of interpretation of other speakers is useful in seeing why this is so. For if we think about our application of the concepts of the attitudes to other agents, it is clear that their utility lies in their enabling us to discern a pattern in the behavior of others that is usefully projectible. As Daniel Dennett (1987 ) puts it, adopting “the intentional stance” enables us to explain and predict the behavior of complicated systems in a way that would otherwise be practically impossible. Thus, contrast trying to predict where someone's body will be tomorrow at noon on the basis of its physical constitution and the laws of nature, on the one hand, with trying to predict where it will be on the basis of his intending to keep a lunch appointment with you tomorrow at the faculty club, on the other. It is the patterns among the attitudes and their relations to behavior imposed by the requirement that agents be largely rational that make for the practical utility of descriptions of objects in terms of propositional attitudes. That is, if a system has successfully been understood as an agent (and, hence, as rational) in the past, so that we have succeeded in interpreting its behavior as the product of rational agency, then, on the assumption that it continues to be an agent, we can usefully (if roughly) predict its future behavior from its present attitudes, any new attitudes it acquires, and what we suppose its reasoning powers to be. (Carl Hempel argued that it is the empirical assumption of rationality that enables us to predict agents' future behavior [Fetzer and Hempel 2001 ]; if being largely rational is essential to agency, the relevant empirical assumption is that the system is an agent. Of course, assumptions about the degree of an agent's rationality can still play a role in empirical explanation. In this regard, see also Davidson 1980 , chap. 14.) end p.355
Language as Necessary for Rationality but Not for Thought The discussion has assumed that rationality is constitutive of thought, and, hence, that it is constitutive of language. But, as noted in the introduction, it has been argued that while language is necessary for rationality, rationality is not necessary for thought. On the face of it, both views cannot be correct. However, if we consider what has been said in favor of the view that language is required for rationality, though rationality is not required for thought, we will see that proponents of the view have in mind stronger requirements on
rationality than those we have imposed above. Thus, these two views, though prima facie in conflict, are not actually so. The view that language is necessary for rationality, but that rationality is not necessary for thought, has an ancient pedigree. Aristotle, and the Stoics, following him, held that while nonlinguistic animals had mental capacities, they did not have the capacity for reason, or, consequently, for belief (Sorabji 1996 ). This view was grounded on two claims. The first is that belief is assent or conviction as a result of persuasion (by others or oneself). The second is that reason is absent where there is no capacity for persuasion. It can be seen, however, that this is not so much a denial of anything we have said above as a use of a stronger conception of “belief” (doxa) or “reason” (logos) than we have been considering. For this is to think of doxa as what we would call belief arrived at on the basis of reflective reasoning, and logos as the capacity for reflective reasoning. This brings out an important point about how “rationality” is now predominantly understood: not as a matter in the first instance of the capacity to engage in reflective reasoning but rather as a matter of thought and action being seen as reasonable in the light of evidence, belief, and desire, whether the agent reflects on these or not. Being rational, in the sense characterized in the previous section, and having the power of rational reflection are not the same. More recently, Jonathan Bennett (1989 , 93) has argued that “possession of language is necessary for rationality” (though not sufficient), on the grounds that (1)“Rationality requires the ability to manifest in behavior judgments about what is particular and past and what is general, that is, to manifest behavior that is appropriate or inappropriate to that which is not both particular and present”; (2)“Only linguistic behavior can be appropriate or inappropriate to that which is not both particular and present” (87). Bennett explicitly denies, however, that “rationality” as it appears in this claim and argument “has anything to do with `rationality' in any contemporary sense of that term” (viii), and is rather intended “to stand for whatever human possession it is that creates a mentalistic difference of kind between us and other terrestrial animals” (vii, 4–5), that is, with giving a sense to “rational” that would truly make us the only rational animals. It is not surprising, given this stipulation, that it should turn out that only linguistic beings are rational. This thesis, then, like Aristotle's, and others in a similar vein, hinges on a conception of rationality that is not the one at issue above in the claim that rationality is necessary for any thought.
4. Language and Thought The final question we take up is the relation between language and thought. If thought is prior to language in the sense that there can be thinking beings who are not linguistic beings, then the rationality of speakers is inherited (in part) from the rationality required
of agents. If language is required for thought, on the other hand, then there is a more intimate connection between having the power of speech and being rational. For then the power of rational thought and the power of speech would be interlocking capacities, each required for the other. The majority view, unsurprisingly, is that thought does not require language. It is easy to see why this is so. First, it is natural and effective to adopt the intentional stance toward many animals. We explain the behavior of both domestic and feral animals by attributing to them beliefs about the world around them and desires similar to ours in basic respects. As Hume says, “Whenwe see other creatures, in millions of instances, perform like actions [to ours], and direct them to like ends, all our principles of reason and probability carry us with an invincible force to believe the existence of like causes” (Hume [1739 ] 1978 , 176). Second, we suppose ourselves to be continuous with the rest of the natural world, and it can seem incredible that human beings should represent the only evolved animals capable of even rudimentary thought. This seems especially incredible in the light of the natural view that language is a means of expressing thoughts that we antecedently possess. This idea is clearly expressed in book 3 of Locke's Essay Concerning Human Understanding ([1690b] 1987), which has had a profound effect on the development of thinking about the mind, even though it has long been recognized that Locke's theory is inadequate in its details. Given the ease with which we can see the nonlinguistic behavior of other animals as expressive of the same sorts of things that similar behavior in us expresses, the view that language merely adds the capacity to express the same basic types of psychological state can seem compelling. There have been dissenters, however, from ancient times to the present, in different forms. Aristotle, as we've seen, held that animals were incapable of belief without language (though he seems to have had in mind more by doxa than we end p.357 would require of belief). Plato held this too, though by virtue of holding implausibly that thought is silent speech; since Plato held that animals have thought, he attributed to them the power of language as well. Descartes famously denied nonlinguistic animals had minds (Descartes 1984a , 140–41; Descartes 1984b , 302–3; Malcolm [1973] 1991 ). Kant thought only the fully rational could have beliefs, and only those who possessed language could be fully rational. In contemporary philosophy, the independence of thought from language has been most famously challenged by Donald Davidson (Davidson 1984 , chap. 11; Davidson 2001 , chap. 7). Davidson gives two arguments for the claim that only speakers have thoughts. The first is the argument from holism (cf. Stich 1979 ). The argument focuses on the kind of evidence available for the attribution of attitudes in the absence of verbal behavior. The Argument from Holism (A1)Beliefs and other attitudes are ascribable only in dense networks of attitudes. (A2)Attributing a dense network of attitudes to an agent requires for support a rich pattern of behavior that gives substance to the attributions.
(A1)Beliefs and other attitudes are ascribable only in dense networks of attitudes. (A3)The pattern of behavior required cannot be sustained in the absence of verbal behavior. (A4)Therefore, only linguistic animals can have propositional attitudes.
Davidson recognizes that this argument is not conclusive. The main problem lies in sustaining premise (A3). The support comes from recognizing that there is something arbitrary about how we choose to describe the supposed attitudes of nonlinguistic animals. As Davidson says, If we really can intelligibly ascribe single beliefs to a dog, we must be able to imagine how we would decide whether the dog has many other beliefs of the kind necessary for making sense of the first. It seems to me that no matter where we start, we very soon come to beliefs such that we have no idea at all how to tell whether a dog has them, and yet such that, without them, our confident first attribution looks shaky. (Davidson 2001 , chap. 7, 98) A very complex pattern of behavior must be observed to justify the attribution of a single thought.I think there is such a pattern only if the agent has language. (100) Thus, what is there to choose between saying that Rover thinks a cat went up that tree or saying that Rover thinks a furred pointy-eared quadruped with whiskers end p.358 scurried up the oak he's barking up? Is there any further behavior that Rover could display that would decide which of these was more appropriate? It seems clear we should not take the particular sentence chosen for the belief attribution too seriously. How then can we take seriously the idea that Rover has any beliefs at all, for if he does have beliefs, must they not have determinate contents? It is undeniable, however, that there is a pattern of activity that is captured equally well for practical purposes by either of our two attributions, but not by attributing to Rover the belief that the cat has vanished or turned into an acorn: Rover barks up the tree, peers into the branches, runs around its base, and finally settles down to wait (as we put it). (And there is the tale of Chrysippus's hunting dog, who, faced with three roads, checks the scent down two, before setting off down the last without checking.) In light of this, perhaps we can characterize Rover's attitudes with classes of sentences that would do equally well for the purposes of accounting for his nonverbal behavior. Each particular sentence in the class attributes too much conceptual sophistication to Rover. To characterize his belief as captured equally well by any of them is, however, to characterize exactly the determinate content of his belief and the degree of refinement of his concepts, though we do not have words to express this in a single sentence. We can press into service Davidson's own favored analogy for disarming the threat of the indeterminacy of interpretation (Davidson 1984 , chap. 10, 154). Like the different scales we could use in measuring temperature, the different belief sentences we use to keep track of Rover's behavior (each of course requiring a supporting pattern of related
ascriptions) all equally well capture the phenomenon we are interested in, which is different from saying that there isn't any (Jeffrey 1985 ). Davidson's primary argument, though, rests on claims about what is required to have beliefs, namely, the concept of belief, and what is, in turn, required to have the concept of belief, namely, language. The basic argument is as follows (Davidson 1984 , chap. 11, 170; Davidson 2001 , chap. 7, 102).
The Argument from the Concept of Belief (B1)One can have propositional attitudes (thoughts) only if one has beliefs. (B2)One can have beliefs only if one has the concept of belief. (B3)One can have the concept of belief only if one has a language. (B4)Therefore, one can have propositional attitudes (thoughts) only if one has a language. The crucial premises are (B2) and (B3). Davidson offers an argument for (B2) in “Rational Animals” (Davidson 2001 , chap. 7, 104). end p.359
The Argument from Surprise (C1)“One cannot have a general stock of beliefs of the sort necessary for having any beliefs at all without being subject to surprises that involve beliefs about the correctness of one's own beliefs. Surprise about some things is a necessary and sufficient condition of thought in general.” (C2)“Surprise requires the concept of belief.” (C3)Therefore, one has beliefs only if one has the concept of belief.
(C2) seems true. Surprise requires recognition that something one thought was so is not, and so requires that one have the concept of belief. (C1) appears more susceptible to challenge, however. Davidson appeals bluntly to intuition, but this seems unconvincing. Child psychologists have argued that children (after initial language acquisition) pass through a developmental stage in which they cannot recognize that they had false beliefs (Gopnik 1990 ; Gopnik and Astington 1988 ; Perner, Leekam, and Wimmer 1987 ). If so, then despite possessing a language, and so thought, they are not capable of being
surprised. Whether this is right or not, it certainly seems possible. But if so, there can be no a priori requirement on thought or language that one be capable of being surprised. (C1) would furthermore lead to the surprising conclusion that there cannot be an omniscient being, since such a being is never surprised. However, we can secure (C3) by another route. For, granting that belief requires agency, having beliefs requires having beliefs about actions serving as means to ends (a belief that doing this is likely to bring about that), and so the concept of action. But the concept of an action involves in turn the concepts of belief, desire, and intention. Thus, one can have beliefs at all only if one has the concept of belief. Let us therefore grant (B2). That leaves (B3). Davidson's main argument for (B3) can be reconstructed as follows (see Davidson 1984 , chap. 11; Davidson 2001 , chap. 3, chap. 7, 103–5, chap. 13, 202 for (D3), chap. 14):
The Argument from the Concept of Error (D1)To have the concept of a belief, one must have the concept of error, or, what is the same thing, of objective truth, the contrast between how things are represented and how they are. (D2)The claim that a creature possesses the concept of error, or objective truth, stands in need of grounding: this must take the form of giving an account of how there could be scope in the creature's experience for correct application of the concept. end p.360 (D3)We can understand how there could be scope for the application of the concept of error in a creature's experience if, and only if, we conceive of it as a creature that is, has been, or is potentially (fixing the creature's capabilities) in communication with other creatures, and so able to use the concept of error as a tool in interpretation to achieve a better rational fit between a speaker's behavior and the beliefs and meanings we attribute to him; that is, the concept of error would have some work to do for interpreters of others' speech (there would be scope for its application in their practice), but not otherwise. (D4)Therefore, from (D2)–(D3), to have the concept of error or objective truth one must be, or have been, or potentially be in communication with others. (D5)Therefore, from (D1) and (D4), to have the concept of belief, one must be, have been in, or potentially be in communication with others. Since communication with others requires one to have a language, (D5) suffices to establish (B3). Davidson nowhere states (D2) explicitly, but it appears to be in the background in the argument, which proceeds by asking after conditions that make sense of there being scope for the application of error in a creature's behavior or experience. Davidson states (D3) in a stronger form than given here, as requiring actual communication, but this seems too strong and is not required for the conclusion.
(D2) may be challenged on the grounds that all that is required for having a concept is that a creature has reason to think that there is scope for correct application of it in his experience. This is a familiar difficulty in transcendental arguments that rely on conditions for the possession of concepts to get to their correct application. However, even granting (D2), (D3) seems questionable. For while communication provides scope for the application of the concept of error, it is not clear why many other activities do not provide just as much scope for its application, such as, for example, correcting one's own past beliefs in the light of new evidence, or explaining behavior of a nonlinguistic animal that otherwise seems irrational in the light of a false belief, for example, explaining Rover's barking up the wrong tree by attributing to him the false belief that that is where the cat is. Davidson's argument for the claim that language is essential to thought is therefore inconclusive. Consequently, we cannot support premise (2) in the following argument for the conclusion that language is necessary for rationality by way of Davidson's argument: (1) a being can think iff it is largely rational; (2) a being can think iff it possesses the power of speech; therefore, (3) a being is largely rational iff it possesses the power of speech. (See Heil 1992 , chap. 5, and Lepore and Ludwig 2004, chap. 22, for further discussion.)
5. Summary Rationality is essential for thought and agency. It is constitutive of the propositional attitudes that they are attitudes of an agent, and nothing is an agent except insofar as its behavior can be interpretable as expressing largely rational behavior and thought. Since agency is a condition on language, rationality is essential for language as well. To say that thought requires rationality is not to say thinkers, and speakers, are perfectly rational, incapable of mistakes, or follies, neuroses, psychoses, and so on. Rather, what mistakes agents make, what irrational behaviors they engage in, are identifiable only against a background of largely rational thought and behavior, which makes sense of the attitudes involved in the irrational behavior or reasoning. Language may be argued to be essential for rationality only by employing a special standard for rationality, requiring advanced reasoning abilities, or by arguing that one can have propositional attitudes only if one is a speaker. Davidson's controversial arguments for the latter position are not conclusive. The ease with which we can explain animal behavior by attributions of propositional attitudes on the basis of the same sort of nonverbal behavior we often use with human beings suggests that thought, agency, and, hence, rationality, are not confined to linguistic beings. Language is not a precondition of rationality, but rather amplifies our powers of reasoning. end p.362
chapter 19 RATIONALITY AND SCIENCE Paul. Thagard
Are scientists rational? What would constitute scientific rationality? In the philosophy of science, these questions are usually discussed in the context of theory choice: What are the appropriate standards for evaluating scientific theories, and do scientists follow them? But there are many kinds of scientific reasoning besides theory choice, such as analyzing experimental data. Moreover, reasoning in science is sometimes practical, for example when scientists decide what research programs to pursue and what experiments to perform. Scientific rationality involves groups as well as individuals, for we can ask whether scientific communities are rational in their collective pursuit of the aims of science. This chapter provides a review and assessment of central aspects of rationality in science. It deals first with the traditional question, What is the nature of the reasoning by which individual scientists accept and reject conflicting hypotheses? I will also discuss the nature of practical reason in science and then turn to the question of the nature of group rationality in science. The remainder of the chapter considers whether scientists are in fact rational, that is, whether they conform to normative standards of individual and group rationality. I consider various psychological and sociological factors that have been taken to undermine the rationality of science. end p.363
What Is Science For? First, however, it is necessary to deal with a prior issue: What are the goals of science? In general, rationality requires reasoning strategies that are effective for accomplishing goals, so discussion of the rationality of science must consider what science is supposed to accomplish. To begin, we can distinguish between the epistemic and the practical goals of science. Possible epistemic goals include truth, explanation, and empirical adequacy. Possible practical goals include increasing human welfare through technological advances. My view is that science has all of these goals, but let us consider some more extreme views. Some philosophers have advocated the view that the primary epistemic aim of science is the achievement of truth and the avoidance of error (Goldman 1999 ). On this view, science is rational to the extent that the beliefs that it accumulates are true, and scientific reasoning is rational to the extent that it tends to produce true beliefs. The philosophical position of scientific realism maintains that science aims for true theories and to some extent accomplishes this aim, producing some theories that are at least approximately true. In contrast, the position of antirealism is that truth is not a concern of science. One of the most prominent antirealists is Bas van Fraassen (1980 ), who argues that science aims only for empirical adequacy: scientific theories should make predictions about observable phenomena but should not be construed as true or false. The antirealist view, however, is at odds with the practice and success of science (see Psillos 1999 for a systematic defense). Most scientists talk and act as if they are trying to figure out how the world actually works, not just attempting to make accurate predictions. Moreover, the impressive technological successes of science are utterly mysterious unless the scientific
theories that made them possible are at least approximately true. For example, my computer would not be processing this chapter unless there really are electrons moving through its silicon chips. But truth is not the only goal of science. The most impressive accomplishments of science are not individual facts or even general laws, but broad theories that explain a great variety of phenomena. For example, in physics the theory of relativity and quantum theory each provide understanding of many phenomena, and in biology the theory of evolution and genetic theory have very broad application. Thus a major part of what scientists strive to do is to generate explanations that tie together many facts that would individually not be very interesting. A scientist who aimed only to accumulate truths and avoid errors would be awash in trivialities. Hence science aims for explanation as well as truth. These two goals subsume the goal of empirical adequacy, because for most scientists the point of describing and predicting observed phenomena is to find out what is true about them and to explain them. But there are also practical goals that science accomplishes. Nineteenthend p.364 century physicists such as Faraday and Maxwell were primarily driven by epistemic goals of understanding electrical and magnetic phenomena, but their work made possible the electronic technologies that now pervade human life. Research on topics such as superconductivity and lasers has operated with both scientific and technological aims. Molecular biology is also a field that began with primarily epistemic aims but that has increasingly been motivated by potential practical applications in medicine and agriculture. Similarly, the major focus of the cognitive sciences such as psychology and neuroscience has been understanding the basic mechanisms of thinking, but there have also been practical motivations such as improving education and the treatment of mental illnesses. It is clear, therefore, that one aim of the scientific enterprise is the improvement of human welfare through technological applications. This is not to say that each scientist must have that aim, since many scientists work far from areas of immediate application, but science as a whole has made and should continue to make technological contributions. More critical views on the practical aims of science are extant. It has been claimed that science functions largely to help maintain the hegemony of dominant political and economic forces by providing ideologies and technologies that forestall the uprising of oppressed peoples. This claim is a gross exaggeration, but there is no question that the products of scientific research can have adverse effects—for example, the use of dubious theories of racial superiority to justify social policies and the use of advanced technology to produce devastating weapons. But saying that the aims of science are truth, explanation, and human welfare does not imply that these aims are always accomplished, only that these are the aims that science generally does and should have. We can now address the question of what strategies of rational thinking best serve the accomplishment of these aims.
Models of Individual Rationality
Consider a recent example of scientific reasoning, the collision theory of dinosaur extinction. Since the discovery of dinosaur fossils in the nineteenth century, scientists have pondered why the dinosaurs became extinct. Dozens of different explanations have been proposed, but in the past two decades one hypothesis has come to be widely accepted: dinosaurs became extinct around 65 million years ago because a large asteroid collided with the earth. Evidence for the collision hypothesis includes the discovery of a layer of iridium (a substance more common in asteroids than on earth) in geological formations laid down around the same end p.365 time that the dinosaurs became extinct. What is the nature of the reasoning that led most paleontologists and geologists to accept the collision hypothesis and reject its competitors? I shall consider three main answers to this question, derived from confirmation theory, Bayesian probability theory, and the theory of explanatory coherence. In each case, I will describe a kind of ideal epistemic agent and consider whether scientists are in fact agents of the specified kind.
Confirmation and Falsification Much work in the philosophy of science has presumed that scientists are confirmation agents that operate roughly as follows (see, e.g., Hempel 1965 ). Scientists start with hypotheses that they use to make predictions about observable phenomena. If experiments or other observations show that the predictions are true, then the hypotheses are said to be confirmed. A hypothesis that has received substantial empirical confirmation can be accepted as true, or at least as empirically adequate. For example, the hypothesis that dinosaurs became extinct because of an asteroid collision should be accepted if it as been confirmed by successful predictions. Popper 1959 argues that scientists should not aim for confirmation but should operate as the following sort of falsification agents. Scientists use hypotheses to make predictions, but their primary aim should be to find evidence that contradicts the predicted results, leading to the rejection of hypotheses rather than their acceptance. Hypotheses that have survived severe attempts to falsify them are said to be corroborated. On this view, the proponents of the collision theory of dinosaur extinction should attempt to falsify their theory by stringent tests and only then consider them as corroborated, but not as accepted as true. Although hypotheses are often used to make predictions, the process of science is much too complex for scientists to function generally as either confirmation agents or falsification agents. In particular, it is exceedingly rare for scientists to set out to refute their own hypotheses, and, given the difficulty of performing complex experiments, it is fortunate that they aim for confirmations rather than refutations. There are many reasons why an experimental prediction might fail, ranging from problems with instruments or
personnel to failure to control for key variables. A falsification agent would frequently end up throwing away good hypotheses. But scientists are not just confirmation agents either, since hypotheses often get support not just from new predictions, but also from explaining data already obtained. Moreover, it often happens in science that there are conflicting hypotheses that are to some extent confirmed by empirical data. As Lakatos 1970 argues, the task then is not just to determine what hypotheses are confirmed, but also what hypotheses are better confirmed than their competitors. Hypothesis assessment is rarely a matter of evaluating a hypothesis with respect to its predictions, but rather requires evaluating competing hypotheses, with the best to be accepted and the others to be rejected. There are both probabilistic and explanatory approaches to such comparative assessment.
Probabilities Carnap and numerous other philosophers of science have attempted to use the resources of probability theory to illuminate scientific reasoning (Carnap 1950 , Howson and Urbach 1989 , Maher 1993 ). Probabilistic agents operate as follows. They assess hypotheses by considering the probability of a hypothesis given the evidence, expressed as the conditional probability P(H/E). The standard tool for calculating such probabilities is Bayes's Theorem, one form of which is: P(H/E) = P(H) * P(E/H) / P(E). This says that the posterior probability of the hypothesis H given the evidence E is calculated by multiplying the prior probability of the hypothesis by the probability of the evidence given the hypothesis, all divided by the probability of the evidence. Intuitively, the theorem is very appealing, with a hypothesis becoming more probable to the extent that it makes improbable evidence more probable. Probabilistic agents look at all the relevant evidence, calculate values for P(E) and P(E/H), take into account some prior value of P(H), and then calculate P(H/E). Of two incompatible hypotheses, probabilistic agents prefer the one with the highest posterior probability. A probabilistic agent would accept the collision theory of dinosaur extinction if its probability given the evidence is higher than the probability of competing theories. Unfortunately, it is not so easy as it sounds for a scientist to be a probabilistic agent. Various philosophers, such as Glymour (1980 ) and Earman (1992 ), have discussed technical problems with applying probability theory to scientific reasoning, but I will mention only what I consider to be the three biggest roadblocks. First, what is the interpretation of probability in P(H/E)? Probability has its clearest interpretation as frequencies in populations of observable events; for example, the probability that a die will turn up a 3 is 1/6, meaning that in a large number of trials there will tend to be 1 event in 6 that turns up a 3. But what meaning can we attach to the probability of dinosaur extinction being caused by an asteroid collision? There is no obvious way to interpret the probability of such causal hypotheses in terms of objective frequencies in specifiable populations.
The alternative interpretation is that such probabilities are degrees of belief, end p.367 but there is substantial evidence that people's thinking does not conform to probability theory (see, e.g., Kahneman, Slovic, and Tversky 1982 ; Tversky 1994 ). One might say that the probability of a hypothesis is an idealized degree of belief, but it is not clear what this means. Degree of belief is sometimes cashed out in terms of betting behavior, but what would it mean to bet on the truth of various theories of dinosaur extinction? The second difficulty in viewing scientists as probabilistic agents is that there are computational problems in calculating probabilities in accord with Bayes's theorem. In general, the problem of calculating probabilities is computationally intractable in the sense that the number of conditional probabilities required increases exponentially with the number of propositions. However, powerful and efficient algorithms have been developed for calculating probabilities in Bayesian networks that make simplifying assumptions about the mutual independence of different propositions (Pearl 1988 ). No one, however, has yet used Bayesian networks to simulate a complex case of scientific reasoning such as debates about dinosaur extinction. In contrast, the next section discusses a computationally feasible account of scientific inference based on explanatory coherence. The third difficulty with probabilistic agents is that they may ignore qualitative factors affecting theory choice. Scientists' arguments suggest that they care not only how much evidence there is for a theory, but also about the variety of the evidence, the simplicity of the theory that accounts for it, and analogies between proposed explanations and other established ones. Perhaps simplicity and analogy could be accounted for in terms of prior probabilities: a simpler theory or one offering analogous explanations would get a higher value for P(H) to be fed into the calculation via Bayes's theorem of the posterior probability P(E/H). But the view of probability as subjective degree of belief leaves it mysterious how people do or should arrive at prior probabilities.
Explanatory Coherence If scientists are not confirmation, falsification, or probabilistic agents, what are they? One answer, which goes back to two nineteenth-century philosophers of science, William Whewell and Charles Peirce, is that they are explanation agents. On this view, what scientists do in theoretical inference is to generate explanations of observed phenomena, and a theory is to be preferred to its competitors if it provides a better explanation of the evidence. Theories are accepted on the basis of an inference to the best explanation. Such inferences are not merely a matter of counting which among competing theories explains more pieces of evidence, but also require assessment in terms of the overall explanatory coherence of each end p.368
hypothesis with respect to a scientist's whole belief system. Factors that go into this assessment for a particular hypothesis include the evidence that it explains, its explanation by higher-level hypotheses, its consistency with background information, its simplicity, and analogies between the explanations offered by the hypothesis and explanations offered by established explanations (Harman 1986 , Lipton 1991 , Thagard 1988 ). The major difficulty with the conception of scientists as explanatory agents is the vagueness of concepts such as explanation, inference to the best explanation, and explanatory coherence. Historically, explanation has been conceptualized as a deductive relation, a probabilistic relation, and a causal relation. The deductive conceptualization of explanation fits well with the confirmation and falsification view of agents: a hypothesis explains a piece of evidence if a description of the evidence follows deductively from the hypothesis. Similarly, the probabilistic conceptualization of explanation fits well with the probabilistic view of agents: a hypothesis explains a piece of evidence if the probability of the evidence given the hypothesis is higher than the probability of the evidence without the hypothesis. Like Salmon (1984 ) and others, I prefer a conceptualization of explanation as the provision of causes: a hypothesis explains a piece of evidence if it provides a cause of what the evidence describes. The causal conceptualization must face the problem of saying what causes are and how causal relations are distinct from deductive and probabilistic ones (see Thagard 1999 , chap. 7). Assuming we know what an explanation is, how can we characterize inference to the best explanation? I have shown how a precise and easily computable notion of explanatory coherence can be applied to many central cases in the history of science (Thagard 1992 ). For example, we can understand why the collision theory of dinosaur extinction has been accepted by many scientists but rejected by others by assessing its explanatory coherence with respect to the evidence available to different scientists (see Thagard 1991 for computer simulations of the dinosaur debate using the program ECHO). I prefer to view scientists as explanation agents rather than as confirmation, falsification, or probabilistic agents because this view fits better with the historical practice of scientists as evident in their writings, as well as with psychological theories that are skeptical about the applicability of deductive and probabilistic reasoning in human thinking. But I acknowledge that the probabilistic agent view is probably the most popular one in contemporary philosophy of science; it has largely absorbed the confirmation agent view by the plausible principle that evidence confirms a hypothesis if and only if the evidence makes the hypothesis more probable—that is, P(H/E) > P(H). It is also possible that scientists are not rational agents of any of these types but rather are reasoners of a very different sort. For example, Mayo 1996 develops a view of scientists as modeling patterns of experimental results that are useful for distinguishing errors. Solomon 2001 end p.369 describes scientists as reaching conclusions based on a wide variety of “decision vectors,” ranging from empirical factors such as salience of data to nonempirical factors such as ideology.
Practical Reason As mentioned in this chapter's introduction, there is much more to scientific rationality than accepting and rejecting hypotheses. Here are some of the important decisions that scientists make in the course of their careers: 1.What general field of study should I enter—for example, should I become a paleontologist or a geologist? 2.Where and with whom should I study? 3.What research topics should I pursue? 4.What experiments should I do? 5.With whom should I collaborate?
When scientists make these decisions, they are obviously acting for more than epistemic reasons, entering a field for more reasons than that it would maximize their stock of truths and explanations. Scientists have personal aims as well as epistemic ones, such as having fun, being successful, living well, becoming famous, and so on. Let us now consider two models of scientists as practical decision makers: scientists as utility agents and scientists as emotional agents. The utility agent view is the familiar one from economics, with an agent performing an action because of a calculation that the action has more expected utility than alternative actions, where expected utility is a function of the utilities and probabilities of different outcomes. This view is consonant with the epistemic view of scientists as probabilistic agents and has many of the same difficulties. When scientists are considering between different research topics, do they have any idea of the relevant probabilities and utilities? Suppose I am a molecular biologist doing genome research and have to decide whether to work with yeast or with worms. I may have hunches about which research program may yield the more interesting results, but it is hard to see how these hunches could be translated into anything as precise as probabilities and utilities. A more realistic view of the decision making of scientists and people in general is that we choose the actions that receive the most positive emotional evaluation based on their coherence with our goals (Thagard 2000 , chap. 6; Thagard 2001 ). On this view, decision making is based on intuition rather than on numerical calculation: unconsciously we balance different actions and different goals, arriv end p.370
ing at a somewhat coherent set of accepted ones. The importance of goals is affected by how they fit with other goals as well as with the different actions that are available to us. We may have little conscious awareness of this balancing process, but the results of the process come to consciousness via emotions. For example, scientists may feel excited by a particular research program and bored or even disgusted by an alternative program. Psychologists use the term valence to refer to positive or negative emotional evaluations. For discussions of the role of emotions in scientific thinking, see Thagard (2002a , b). Like Nussbaum (2001 ), I view emotions as intelligent reactions to perceptions of value, including epistemic value. Just as there is a concordance between the probabilistic view of epistemic agents and the utility view of practical agents, there is a concordance between the explanatory coherence view of epistemic agents and the emotional coherence view of practical agents. In fact, emotions play a significant role in inference to hypotheses as well as in inference to actions, because the inputs to and outputs from both kinds of inference are emotional as well as cognitive. The similarity of outputs is evident when scientists appreciate the great explanatory power of a theory and characterize it as elegant, exciting, or even beautiful. As with practical judgments of emotional coherence in practical decision making, we have no direct conscious access to the cognitive processes by which we judge some hypotheses to be more coherent than others. What emerges to consciousness from a judgment of explanatory coherence is often emotional, in the form of liking or even joy with respect to one hypothesis, and dislike or even contempt for rejected competing hypotheses. For example, when Walter and Luis Alvarez came up with the hypothesis that dinosaurs had become extinct because of an asteroid collision, they found the hypothesis not only plausible but even exciting (Alvarez 1998 ). In contrast, some skeptical paleontologists thought the hypothesis was not only dubious but also ridiculous. Emotional inputs to hypothesis evaluation include the varying attitudes that scientists hold toward different experimental results and even for different experiments—any good scientist knows some experiments are better than others. Another kind of emotional input is analogical: a theory analogous to a positively viewed theory such as evolution will have greater positive valence than one that is analogous to a scorned theory such as cold fusion. Thus my view of scientists as explanatory-emotional agents is very different from the view of them as probabilistic-utility agents. My emphasis on emotions will probably have readers wondering whether scientists are rational at all. Perhaps they are just swayed by their various intellectual prejudices and personal desires to plan research programs and accept hypotheses in ways that disregard the epistemic aims of truth and explanation. There are, unfortunately, cases where scientists are deviant in these ways, with disastrous results such as fraud and other kinds of unethical behavior. But the temperaments and training of most scientists are such that they have an emotional attachment to the crucial epistemic aims. Many scientists become scientists because they enjoy finding out how things work, so that the aims of truth and explanation are with them from the beginnings of their scientific training. These attachments can be fostered by working with advisors who not only value these aims but also transmit their emotional evaluations to the students and postdoctoral fellows with whom they work. So, for most scientists, a commitment to fostering explanation and truth is an emotional input into their practical decision making.
Models of Group Rationality As Kuhn (1970 ) and many other historians, philosophers, and sociologists of science have noted, science is not merely a matter of individual rationality. Scientists do their work in the context of groups of various sizes, from the research teams in their own laboratories to the community of scientists working on similar projects to the overall scientific community. As I have documented elsewhere (Thagard 1999 , chap. 11), most scientific articles have multiple authors, and the trend is toward increasing collaboration. In addition, all scientists operate within the context of a wider community with shared societies, journals, and conferences. Therefore the question of the rationality of science can be raised for groups as well as individuals: What is it for a group of scientists to be collectively rational, and are such groups generally rational? I will assume that groups of scientists have the same primary aims that I attributed to science in general: truth, explanation, and human welfare via technological applications. It might seem that the rationality of scientific groups is just the sum of the rationality of the individuals that comprise them. Then a group is rational if and only if the individual scientists in it are rational. But it is possible to have individual rationality without group rationality, if the pursuit of scientific aims by each scientist does not add up to optimal group performance. For example, suppose that each scientist rationally chooses to pursue exactly the same research strategy as the others, with the result that there is little diversity in the resulting investigations, and paths that would be more fertile with respect to truth and explanation are not taken. Philosophers such as Kitcher (1993 ) have emphasized the need for cognitive diversity in science. On the other hand, it might be possible to have group rationality despite lack of individual rationality. Hull (1989 ) has suggested that individual scientists who seek fame and power rather than truth and explanation may in fact contribute to the overall aims of science, because their individualistic pursuit of nonepistemic end p.372 motives in fact leads the scientific group as a whole to prosper. This is analogous to Adam Smith's economic model in which individual greed leads to overall economic growth and efficiency. It is important to recognize also that group rationality in science is both epistemic and practical. Of a particular scientific community, we can ask two kinds of question: (1)Epistemic: Given the evidence, what should be the distribution of beliefs in the community? (2)Practical: What should be the distribution of research initiatives in the community?
For the epistemic question, it might be argued that if all scientists have access to the same evidence and hypotheses, then they should all acquire the same beliefs. Such unanimity would, however, be detrimental to the long-term success of science, since it would reduce cognitive diversity. For example, if Copernicus had been enmeshed within the Ptolemaic theory of the universe, he might never have generated his alternative heliocentric theory, which turned out to be superior with respect to both truth and explanation. Similarly, in the dinosaur case Walter Alvarez would never have formulated his theory of why dinosaurs became extinct if he had been a conventional paleontologist. Moreover, epistemic uniformity would contribute to practical uniformity, which would clearly be disastrous. It would be folly to have all scientists within a scientific community following just a few promising leads, since this would reduce the total accomplishment of explanations as well as retard the development of novel explanations. Garrett Hardin (1968 ) coined the term “tragedy of the commons” to describe a situation in which individual rationality could promote group irrationality. Consider sheep herders who share a common grazing area. Each herder separately may reason that adding one more sheep to his or her herd would not have any serious effect on the common area. But such individual decisions might collectively produce overgrazing, so that there is not enough food for any of the sheep, with the result that all sheep herders are worse off. Analogously, we can imagine in science and other organizations a kind of “tragedy of consensus,” in which the individuals all reach similar conclusions about what to believe, stifling creative growth. So, what should be our model of group rationality in science? Kitcher 1993 and Goldman 1999 develop models of group rationality that assume that individual scientists are probabilistic agents. Although these analyses are interesting with respect to cognitive diversity and truth attainment, I do not find them plausible because of the problems with the probabilistic view discussed in the last section. As an alternative, I have developed a model of scientific consensus based on explanatory coherence. This model is called CCC, for consensus = coherence + communication (Tha end p.373 gard 2000 , chap. 10). It assumes that each scientist is an explanation agent, accepting and rejecting hypotheses on the basis of their explanatory coherence with evidence and alternative hypotheses. Communication takes place as the result of meetings between scientists in which they exchange information about available evidence and hypotheses. If all scientists acquire exactly the same information, then they will agree about what hypotheses to accept and reject. However, in any scientific community, exchange of information is not perfect, so that some scientists may not hear about some of the evidence and hypotheses. Moreover, different scientists have different antecedent belief systems, so the overall coherence of a new hypothesis may be different for different scientists. Ideally, however, if communication continues there will eventually be community consensus as scientists accumulate the same sets of evidence and hypotheses and therefore reach the same coherence judgments. The CCC model has been implemented as a computational extension of the explanatory coherence program ECHO, in which individual scientists evaluate hypotheses on the basis of their explanatory coherence but also exchange hypotheses and evidence with other scientists. These
simulated meetings can be either pairwise exchanges between randomly selected pairs of scientists or “lectures” of the sort that take place at scientific conferences in which one scientist can broadcast sets of hypotheses and evidence to a group of scientists. Of course, communication is never perfect, so it can take many meetings before all scientists acquire approximately the same hypotheses and evidence. I have performed computational experiments in which different numbers of simulated scientists with varying communication rates achieve consensus in two interesting historical cases: theories of the causes of ulcers, and theories of the origins of the moon. The CCC model shows how epistemic group rationality can arise in explanation agents who communicate with each other, but it tells us nothing about practical group rationality in science. One possibility would be to attempt to extend the probabilistic-utility model of individual practical reason. On this model, each scientist makes practical decisions about research strategy based on calculations concerning the expected utility of different courses of action. Research diversity arises because different scientists attach different utilities to various experimental and theoretical projects. For reasons already given, I would prefer to extend the explanatory-emotional model described in the previous section. The extension arises naturally from the CCC model just described, except that in large, diverse communities we should not expect the same degree of practical consensus as there is of epistemic consensus, for reasons given below. For the moment, let us focus on particular research groups rather than on whole scientific communities. At this level, we can find a kind of local consensus that arises because of emotional coherence and communication. The characteristics of the group include the following: end p.374 1.Each scientist is an explanation agent with evidence, hypotheses, and the ability to accept and reject them on the basis of explanatory coherence. 2.In addition, each scientist is an emotional agent with actions, goals, valences, and the ability to make decisions on the basis of emotional coherence. 3.Each scientist can communicate evidence and hypotheses with other scientists. 4.Each scientist can, at least sometimes, communicate actions, goals, and valences to other scientists. 5.As the result of cognitive and emotional communication, consensus is sometimes reached about what to believe and also about what to do. The hard part to implement is the component of (4) that involves valences. It is easy to extend the CCC model of consensus to include emotional coherence simply by allowing actions, goals, and valences to be exchanged just like evidence, hypotheses, and explanations. In real life, valences are not so easily exchanged as verbal information about actions, goals, and what actions accomplish which goals. Just hearing someone say that they really care about something does not suffice to make you care about it too, nor should it, because your goals and valences may be orthogonal or even antagonistic to mine. So in a
computational model of emotional consensus, the likelihood of exchange of goals and valences in any meeting should be much lower than the likelihood of exchange of hypotheses, evidence, and actions. Still, in real-life decision making involving scientists and other groups such as corporate executives, emotional consensus is sometimes reached. What are the mechanisms of valence exchange—that is, how do people pass their emotional values on to other people? Two relevant social mechanisms are emotional contagion and attachment-based learning. Emotional contagion occurs when person A expresses an emotion and person B unconsciously mimics A's facial and bodily expressions and then begins to acquire the same emotion (Hatfield, Cacioppo, and Rapson 1994 ). For example, if a group member enthusiastically presents a research strategy, then the enthusiasm may be conveyed through both cognitive and emotional means to other members of the group. The cognitive part is that the other group members become aware of possible actions and their potential good consequences, and the emotional part is conveyed by the facial expressions and gestures of the enthusiast, so that the positive valence felt by one person spreads to the whole group. Negative valence can also spread, not just from a critic pointing out drawbacks to a proposed action as well as more promising alternatives, but also by contagion of the negative facial and bodily expressions. Another social mechanism for valence exchange is what Minsky (2001 ) calls attachment-based learning. Minsky points out that cognitive science has developed good theories of how people use goals to generate subgoals, but has had little to end p.375 say about how people acquire their basic goals. Similarly, economists employing the expected-utility model of decision making take preferences as given, just as many philosophers who hold a belief-desire model of rationality take desires as given. Minsky suggests that basic goals arise in children as the result of praise from people to whom the children are emotionally attached. For example, when young children share their toys with their playmates, they often receive praise from their parents or other caregivers. The parents have positive valence for the act of sharing, and the children may also acquire a positive emotional attitude toward sharing as the result of seeing that it is something cared about by people whom they care about and who care about them. It is not just that sharing becomes a subgoal to accomplish the goal of getting praised by parents; rather, being kind to playmates becomes an internalized goal that has intrinsic emotional value to the children. I conjecture that attachment-based learning also occurs in science and other contexts of group decision making. If your supervisor is not just a boss but also a mentor, then you may form an emotional attachment that makes you particularly responsive to what the supervisor praises and criticizes. This makes possible the attachment-based transmission of positive values such as zeal for truth and understanding, or, more locally, for integrity in dealing with data and explanations. Notice that both emotional contagion and attachment-based learning require quite intense interpersonal contacts that will not be achieved in a large lecture hall or video conference room, let alone through reading a published article. The distinguished social psychologist, Richard Nisbett, told me that he learned how to do good experiments through discussions
with his supervisor, Stanley Schacter. Nisbett said, “He let me know how good my idea was by grunts: noncommittal (`hmmm'), clearly disapproving (`ahnn') or (very rarely) approving (`ah!').” These grunts and their attendant facial expressions conveyed emotional information that shaped the valences of the budding researcher. Accordingly, when I extend my CCC model of consensus as coherence plus communication to include group decisions, I will include two new variables to determine the degree of valence transmission between agents: degree of personal contact, and degree of attachment. If personal contact and attachment are high, then the likelihood of valence transmission will be much greater than in the ordinary case of scientific communication, in which the success of verbal transmission of information of hypotheses, evidence, and actions is much higher than the transmission of valences. There may, however, be quasi-verbal mechanisms for valence transfer. Thagard and Shelley (2001 ) discuss emotional analogies whose purpose is to transfer valences as well as verbal information. For example, if a scientist presents a research project as analogous to a scientific triumph such as the asteroid theory of dinosaur extinction, then listeners may transfer the positive value they feel for the asteroid theory to the proposed research project. Alternatively, if a project is analogous to the cold fusion debacle, then the negative valence attached to that case may be projected onto the proposed project. Thus emotional analogies are a third mechanism, in addition to emotional contagion and attachment-based learning, for transfer of valences. All three mechanisms may interact with each other, for example when a mentor uses an emotional analogy and facial expressions to convey values to a protégé. Alternatively, the mentor may function as a role model, providing a different kind of emotional analogy: students who see themselves as analogous to their role models may tend to transfer to themselves some of the motivational and emotional characteristics of their models. I hope it is obvious from my discussion of practical group rationality in science why science need not succumb to the tragedy of consensus, especially with respect to practical rationality. Communication between scientists is imperfect, both with respect to cognitive information such as hypotheses and evidence and especially with respect to emotional valences for particular approaches. Scientists may get together for consensus conferences such as the ones sponsored by the National Institutes of Health that regularly deal with controversial issues in medical treatment (see Thagard 1999 , chap. 12, for a discussion). But not all scientists in a community attend such conferences or read the publications that emanate from them. Moreover, the kinds of close interpersonal contact needed for communication of values by emotional contagion and attachment-based learning occur only in small subsets of the whole scientific community. Hence accomplishment of the general scientific aims of truth, explanation, and technological applications need not be hindered in a scientific community by a dearth of practical diversity. Solomon 2001 provides a rich discussion of consensus and dissent in science.
Is Science Rational? A person or group is rational to the extent that its practices enable it to accomplish its legitimate goals. At the beginning of this chapter, I argued that the legitimate goals of
science are truth, explanation, and technologies that promote human welfare. Do scientific individuals and groups function in ways that further these goals, or do they actually pursue other personal and social aims that are orthogonal or even antagonistic to the legitimate goals? I will now consider several psychological and sociological challenges to the rationality of science. Psychological challenges can be based on either cold cognition, which involves processes such as problem solving and reasoning, or hot cognition, which includes end p.377 emotional factors such as motivation. The cold-cognition challenge to scientific rationality would be that people's cognitive processes are such that it is difficult or impossible for them to reason in ways that promote the aims of science. If scientific rationality required people to be falsification agents or probabilistic agents, then the coldcognition challenge would be a serious threat. I cited earlier some of the experimental and historical data that suggest that probabilistic reasoning and falsification are not natural aspects of human thinking. In contrast, there is evidence that people can use explanatory coherence successfully in social judgments (see Read and Marcus-Newhall 1993 ). One might argue that there is evidence that people are confirmation agents, and not very good ones in that they tend toward confirmation bias in looking excessively to confirm their hypotheses rather than falsify them (see Klayman and Ha 1987 ). However, the psychological experiments that find confirmation biases involve reasoning tasks that are much simpler than those performed by actual scientists. Typically, nonscientific subjects are asked to form generalizations from observable data, for example in seeing patterns in numerical sequences. The generalization tasks of real scientists are more complex, in that data interpretation requires determining whether apparent patterns in the data are real or just artifacts of the experimental design. If scientists did not try hard to get their experiments to confirm their hypotheses, the experiments would rarely turn out to be interesting. Notably, trying hard to confirm is not always sufficient to produce confirming results, so scientists sometimes have falsification thrust upon them. But their bias toward finding confirmations is not inherently destructive to scientific rationality. A more serious challenge to the rationality of science comes from hot cognition. Like all people, scientists are emotional beings, and their emotions may lead to distortions in their scientific works if they are attached to values that are inimical to the legitimate aims of science. Here are some kinds of cases where emotions have distorted scientific practice: 1.Scientists sometimes advance their own careers by fabricating or distorting data in order to support their own hypotheses. In such cases, they have greater motivation to enhance their own careers than to pursue truth, explanation, or welfare. 2.Scientists sometimes block the publication of theories that challenge their own by fabricating problems with submitted articles or grant proposals that they have been asked to review. 3.Without being fraudulent or intentionally evil, scientists sometimes unintentionally deceive themselves into thinking that their hypotheses and data are better than those of
1.Scientists sometimes advance their own careers by fabricating or distorting data in order to support their own hypotheses. In such cases, they have greater motivation to enhance their own careers than to pursue truth, explanation, or welfare. their rivals. 4.Scientists sometimes further their careers by going along with politically mandated views—for example, the Nazi rejection of Einsteinian physics and the Soviet advocacy of Lysenko's genetic theories. end p.378
Cases like these show indubitably that science is not always rational. Some sociologists such as Latour (1987 ) have depicted scientists as largely concerned with gaining power through the mobilization of allies and resources. It is important to recognize, however, that the natural emotionality of scientists is not in itself a cause of irrationality. As I documented elsewhere, scientists are often motivated by emotions that further the goals of science, such as curiosity, the joy of discovery, and appreciation of the beauty of highly coherent theories (Thagard, 2002b ). Given the modest incentive structure of science, a passion for finding things out is a much more powerful motivator of the intense work required for scientific success than are extrinsic rewards such as money and fame. Thus hot cognition can promote scientific rationality, not just deviations from it. The mobilization of resources and allies can be in the direct or indirect service of the aims of science, not just the personal aims of individual scientists. A useful response to the question “Is science rational?” is “Compared to what?” Are scientists as individuals more adept than nonscientists at fostering truth, explanation, and human welfare? The history of science and technology over the past two hundred years strongly suggests that the answer is yes. We have acquired very broadly explanatory theories such as electromagnetism, relativity, quantum theory, evolution, germ theory, and genetics. Thousands of scientific journals constitute an astonishing accumulation of truths that ordinary life would never have allowed. Moreover, technologies such as electronics and pharmaceuticals have enriched and lengthened human lives. So the occasional irrationality of individual scientists and groups is compatible with an overall judgment that science is in general a highly rational enterprise. In recent decades, the most aggressive challenge to the ideal of scientists as rational agents has come from sociologists and historians who claim that scientific knowledge is “socially constructed.” Obviously, the development of scientific knowledge is a social as well as an individual process, but the social construction thesis is usually intended to make the much stronger claim that truth and rationality have nothing to do with the development of science. My own view is that an integrated psychological/sociological view of the development of scientific knowledge is perfectly compatible with scientific rationality involving the frequently successful pursuit of truth, explanation, and human welfare (Thagard 1999 ). Crucially, however, the assessment of scientific rationality needs to employ models of individual reasoning and group practices that reflect the thought processes and methodologies of real scientists. Models based on formal logic and probability theory have tended to be so remote from scientific practice that they encourage the inference that
scientists are irrational. In contrast, psychologically realistic models based on explanatory and emotional coherence, along with socially realistic models of consensus, can help to illuminate the often impressive rationality of the enterprise of science. end p.379
chapter 20 ECONOMIC RATIONALITY Paul. Weirich According to an introductory economics textbook by Case and Fair (1992 , 5), “Economics is the study of how individuals and societies choose to use the scarce resources that nature and previous generations have provided.” Economic theory assumes as a first approximation that individuals and societies choose ration-ally. Consequently, it carefully formulates principles of rational decision making. I examine conceptions of rationality common in economics. Although economists treat rationality from multiple perspectives, I describe only a few central views. Other handbook entries supplement my sketch. The title “Economic Rationality” is not meant to suggest that rationality comes in various types, one of which is appropriate in economics. The familiar concept of rationality, shared by all the disciplines, guides principles of rationality in economics. Economics divides into micro- and macroeconomics. The former treats the interaction of consumers and producers, such as price setting, and the latter treats the aggregate results of that interaction, such as the national inflation rate. Microeconomics formulates principles of rational choice. It studies both individual and group decision making. The division between these types of decision is flexible, however. A household or a firm composed of individuals may be treated as a single individual. Economists focus on rationality in choice and action. Rationality in preference and belief arise as ancillary topics. They treat instrumental rationality—that is, the rational adoption of means to reach ends. They also propose standards of end p.380 rationality demanding various types of consistency. For example, Raiffa (1968 , 76), Pindyck and Rubinfeld (1989 , 59), and Case and Fair (1992 , 184) assume that rationality requires the consistency of a preference ranking, understood as the transitivity of preferences and of attitudes of indifference. Consistency of this type goes beyond logical consistency. A household has diverse goals. Besides food, clothing, and shelter, it may want education and entertainment. A firm typically has more limited goals such as profit and reduction of risk. Economists tend to leave to philosophers the formulation of standards of rationality for basic goals, but they assume that the usual goals of households and firms are rational. Economists may seem to condemn as irrational certain basic goals. They commonly warn against the irrationality of being swayed by sunk costs, neglecting opportunity costs, and having pure time preferences. A textbook may, for example, admonish the owner of a languishing restaurant for sinking more money into his business, trying to keep it afloat
against the odds, instead of closing it and accepting the loss of his initial investment. It may criticize a married couple who continue to occupy a large house after their children have grown up and moved away, ignoring the opportunity to sell their house and use the proceeds to buy both another, smaller house and a retirement annuity. It may reprimand an impatient person who picks the lesser of two goods just because it is available a little earlier. One's impression may be that the proscribed behavior is deemed irrational because it proceeds from irrational basic goals. For instance, it may seem that counting sunk costs is deemed irrational because it proceeds from an irrational basic goal never to accept a loss. However, economists typically hold that such mistakes stem from misperception of options and consequences, or muddled thinking. They count the mistakes as errors of judgment rather than as errors in basic goals and so do not regard their warnings as criticisms of basic goals.
1. Conceptions of Rationality in Economics Three conceptions of rationality dominate recent economic theorizing. They are closely related and may be thought of as links in an evolutionary chain rather than as rivals for the same environmental niche. Maximizing Self-Interest A traditional view in economics takes a person's basic goal to be self-interest and takes rationality to be the promotion of self-interest. When several courses of action are open, rationality recommends the one that best promotes self-interest. It says to maximize selfinterest. When the action that maximizes self-interest is uncertain, rationality recommends probability as a guide. Maximize expected self-interest, it says. Edgeworth clearly enunciated the assumption of self-interest. According to him, “The first principle of economics is that every agent is actuated only by self-interest” (1881 , 16). Philosophers call this view psychological egoism. A closely related view is egoism about rationality, the view that rationality requires a person to promote her own interest exclusively. Some presentations of the egoistic view of rationality use technical terminology. Instead of speaking of an agent's interest, they speak of utility for her. Pindyck and Rubinfeld (1989 , 81, 87) define utility for an agent as satisfaction for her and recommend that an agent maximize utility for herself, that is, maximize her satisfaction. Their analyses assume that a firm seeks long-run profits and that rationality requires its acting to maximize long-run profits (246–47, 458). Case and Fair (1992 , 88, 160, 181, 476–77) take utility as happiness, satisfaction, or well-being and say that an agent maximizes, and ought to maximize, utility for herself. Although psychological egoism and egoism are common in economics, they may be adopted as simplifying assumptions met approximately in typical economic transactions. Economists point out that, contrary to common belief, self-interest does not generate a war of all against all. One of the lessons of Adam Smith's The Wealth of Nations is that agents pursuing self-interest are often led to cooperate because cooperation is frequently an excellent means of promoting their own interests. Self-interested agents may trade goods, for instance, because each benefits from the exchange (Smith 1976 , 26–27). Smith proclaims that self-interest leads agents to act as if an invisible hand were guiding
their behavior toward the common good (1976 , 456). In the same vein, as Pindyck and Rubinfeld (1989 , 570) explain, contemporary welfare economics shows that if everyone trades in the marketplace to maximize her satisfaction, and all mutually beneficial trades are completed, the resulting allocation will be efficient in the sense that no alternative allocation yields gains for some without losses for others. Efficiency in this sense is called Pareto optimality after the nineteenth-century Italian economist Vilfredo Pareto. Self-interest has different meanings for different theorists. The standard interpretation takes a person's interest as her welfare or well-being, that is, what is good for her. However, those holding that all people are exclusively self-interested sometimes take a person's interest as satisfaction of her desires, whatever their content. Fulfilling any of her desires counts as promoting her interest. Then the end p.382 claim that people are self-interested says no more than that people pursue their desires. As Sen (1977 , 322–24) observes, it does not rule out altruism, since an altruist pursues her desire to help others. Binmore (1994 , 6, 15, 18–19, 21) claims that rationality amounts to promoting enlightened self-interest, or broad self-interest. He means promotion of one's goals whatever they may be, even if they include showing sympathy toward others and honoring commitments (21, 28). Taking self-interest broadly makes promotion of selfinterest generate the next conception of rationality.
Maximizing Utility During the twentieth century, economists discovered techniques for using an agent's preferences among acts to derive his assignment of probabilities and utilities to the acts' possible outcomes and thereby derive the acts' utilities. Von Neumann and Morgenstern (1953 ) developed and promulgated such methods, but Ramsey (1931 , chap. 7) had worked out similar methods earlier (see Joyce, chap. 8, and Dreier, chap. 9, this volume). The methods derive subjective probabilities and utilities—roughly, degrees of belief and degrees of desire, respectively. They enrich accounts of decision making in the face of risk. They also provide a framework for a broad account of an agent's basic goals. The new utility theory does not restrict the content of those goals. According to Ramsey, “The theory I propose to adopt is that we seek things which we want, which may be our own or other people's pleasure, or anything else whatever, and our actions are such as we think most likely to realize these goods” (1931, 173). Modern utility theory generalizes maximization of self-interest to maximization of personal utility. It takes utility broadly enough to accommodate all goals, not just self-interest. Von Neumann and Morgenstern (1953 , 1, 8–9) say economic tradition characterizes rational behavior as the maximization of utility taken as satisfaction in the case of the consumer and profit in the case of the entrepreneur. However, the utility theory they articulate (chap. 3) does not restrict an agent's goals. Savage (1972 ), who extended von Neumann and Morgenstern's approach to utility, says that the kind of utility they
introduce is dependent on probability in contrast with earlier kinds—for example, satisfaction—that are independent of probability. Their probability-dependent type of utility figures in the representation of preferences among acts involving risk and, he says, is better able to explain rational behavior (1972, 91–104). It does not assume that an agent has selfish goals; it just arithmetizes the relation of preference among acts (1972, 69). Von Neumann and Morgenstern make utility dependent on objective probability, whereas, for greater explanatory power, Savage makes utility dependent on subjective probability. end p.383 Although as Arrow (1963 , 3) observes, economic tradition takes rationality to be some sort of maximization, modern utility theory in taking rationality as maximization of utility departs from the view that rationality is maximization of self-interest. An agent's utility assignment need not follow self-interest. Other goals, including concern for other people, may influence an agent's utility assignment. According to the new view, rationality recognizes all of an agent's goals. Harsanyi (1977 , 10), working within the same school, recognizes a utility function that assigns positive utility to unselfish values. He says that rational agents pursue both the selfish and unselfish values to which their utility functions assign positive utility. Sometimes in modern utility theory maximizing utility goes by other names. Maximizing utility among risky options may be called maximizing expected utility since it depends on probabilities and expectations concerning options' possible outcomes. Maximizing utility and maximizing expected utility are equivalent, however. Utility, being subjective, is relative to information. Because an agent's utility assignment to a risky option depends on the agent's goals and information, changes in the agent's information influence the option's utility assignment. The more evidence the option will realize the agent's goals, the more attractive it is to him. An option's utility and expected utility therefore agree. Sometimes utility is taken to be relative to full information. Under this interpretation, utility may differ from expected utility. Expected utility is relative to actual information, not necessarily full information. Since utility relative to full information may differ from utility relative to actual information, it may differ from expected utility even though utility relative to actual information agrees with expected utility. I treat maximizing utility with respect to actual information, which is the same as maximizing expected utility, and so do not distinguish between maximizing utility and maximizing expected utility. Because utility represents preference, maximizing utility among options is the same as adopting an option at the top of one's preference ranking of options. Thus, Varian's (1984 , 115) view that rationality requires preference maximization agrees with the view that it requires utility maximization. The difference between preference maximization and utility maximization is only terminological. Because modern utility theory derives utilities from preferences, it often bills utility maximization as a form of consistency among preferences, or among choices, or among preferences and choices. Luce and Raiffa (1957 , 31–32) say that it takes utility as a representation of preferences or a guide to consistent action. Barry (1965 , 4–5) says that economists hold that rationality does not require maximizing utility, thought of as
satisfaction, but only requires a consistent pattern of choices. Arrow (1967 , 5) says that rationality is a matter of making choices derivable from a preference ordering. This condition imposes a type of consistency on choices. Harsanyi (1977 , 8) also says that choice behavior is rational if it satisfies certain consistency requirements. According to Broome (1991 , 90–92), modern utility theory claims only that a preference ranking, if rational, may be represented end p.384 by a utility assignment according to which an act's utility equals the expected utility of its possible outcomes. This condition imposes a type of consistency on a preference ranking. Binmore (1994 , 21, 27) says that rationality requires only being consistent in seeking one's goals, which demands consistency between preferences and acts. The train of thought that construes utility maximization as a form of consistency is roughly the following. The techniques for deriving probabilities and utilities from preference rankings assume that preference rankings are transitive and meet other requirements that may be taken as requirements of consistency in a broad sense. In general, the techniques require preference rankings to be as if maximizing utility; preferences among options have to follow their utilities computed with respect to the derived probability and utility assignments for the options' possible outcomes. This global requirement may also be taken as a consistency requirement. It yields utility maximization given utility's derivation from preference rankings. Operationists go further. They define probability and utility in terms of preferences, and some, such as Binmore (1994 , 169), moreover define preferences in terms of choices. As Binmore notes, since operationist definitions of preferences and utilities make them deducible from choices, satisfying preferences or maximizing utility is just consistent choice behavior (1994 , 50–51). For example, consider an operational definition of utility as a mathematical representation of preferences according to which preferences follow utilities. Because the definition entails that preferences follow utilities, it makes that traditional consistency requirement a definitional truth and deprives it of normative force. Similarly, a definition of preference in terms of choice makes the traditional consistency requirement that choice follow preference a definitional truth without normative force. That choices maximize utility becomes a definitional truth. After operationalization, the only consistency requirements with normative force concern consistency of preferences or consistency of choices. See Blackburn 1998 , chap. 6, and Hubin 2001 , sec. 3.
Bounded Rationality The previous sections examined conceptions of rationality requiring maximization of some sort. However, some theorists doubt that humans are capable of the kind of maximization demanded. Simon (1982 ) advocates standards of rationality attuned to human limits, especially cognitive limits. His main proposal starts with the observation that people do not survey all their options at the start of a decision problem. They search
for options during deliberations and entertain them sequentially. Also, to simplify evaluation of an option they just classify it as satis end p.385 factory or not satisfactory, depending on whether it meets their aspirations. To approximate optimization in a practical way, Simon proposes adopting the first satisfactory option discovered (1982, 250–51). Someone selling a house, for example, may rationally accept the first satisfactory offer instead of holding out for an optimal offer. Someone playing chess may adopt the first satisfactory move spotted instead of attempting the Herculean task of identifying the optimal move. Simon calls his procedure “satisficing” and calls theories incorporating such procedures theories of “bounded rationality.” Simon argues that in view of our limited time and cognitive capacity for making decisions, we should not be held to the classical standard of utility maximization. In general, he advocates replacing substantive rules of rationality, such as the rule to maximize utility, with procedural rules of rationality, such as the rule to satisfice (1982, 424–43). Procedural rules prescribe methods of making decisions rather than decisions to be made. They focus on the process rather than the result. They consider how decisions should be made rather than what decisions should be made. Some commonsense rules for making decisions are procedural. The advice to sleep on an important decision before finalizing it, for instance, attends to method rather than result. According to Simon, whereas substantive rationality targets best solutions, procedural rationality targets solutions good in view of human limits. Procedural rationality is practical in the face of problems that are intractable from the perspective of substantive rationality (1982, 428, 431). 1 The rule to satisfice is not purely procedural, since it aims for a satisfactory option. But aspiration levels adjust to the difficulty of finding a satisfactory option, and the search procedure determines which satisfactory option emerges first. So procedure plays a major role in satisficing. Also, the motivation for satisficing is practical. Hence Simon classifies the rule to satisfice as a procedural rule. Many decision theorists have an interest in modifying standards of rationality to better suit the cognitive abilities of humans. Skyrms (1990 , 2–3) advocates using Simon's theory of bounded rationality and a dynamic approach to decision making to articulate the connection between decision rules and game theory. Rubinstein (1998 , 4) advocates using models of bounded rationality, applied especially to human inferential activity, to revise the accounts of equilibrium game theory advances. (See Bicchieri, chap. 10, this volume, on game theory.)
2. Criticisms Because economics concentrates on instrumental rationality, a common complaint is that its conceptions of rationality are too narrow. Irrational goals escape criticism. For example, if an insane agent is bent on self-mutilation, economic rationality may seem not
to block his path. Furthermore, economics' approach to rationality may seem cold and calculating, and thus neglectful of the value of emotion and spontaneity. The charge that economics examines only instrumental rationality is an exaggeration. The view that one should promote self-interest assumes that rationality requires that goal, for instance. Also, economists advance certain social goals as requirements of rationality. Case and Fair (1992 , 21–23) list efficiency, equity, growth, and stability. These goals are assumed in the branch of economics known as economic policy, which is allied with social philosophy. Still, it is true that economics downplays the evaluation of basic goals. It may easily correct the impression that it is averse to their evaluation, however, by acknowledging the need to supplement its principles of instrumental rationality with additional principles of rationality for assessing basic goals. Moreover, economics may acknowledge the value of emotion and spontaneity without abandoning its approach to instrumental rationality. Its principles of instrumental rationality may be formulated to accommodate all values, not just the materialistic ones common in textbook examples. Because economics may respond to the criticisms mentioned without significantly changing its conceptions of rationality, I do not examine those criticisms further. Also, I put aside criticisms of economic rationality prompted by its extension from single agents to groups of agents. Arrow's Possibility Theorem (1963 , chap. 5) presents a problem for this extension. It establishes that there is no satisfactory general procedure for aggregating the rational preference rankings of a group's members into a rational preference ranking for the group. Although some theorists suggest revising standards of rationality for individuals to facilitate their extension to groups, it may turn out that the standards appropriate for individuals just differ from the standards appropriate for groups. Perhaps an aggregation of individual preferences by a market or voting mechanism just does not produce results rational in any collective sense more rich than efficiency. Given this possibility, the problem of extending standards of individual rationality to groups does not yield straightforward criticisms of those standards. The following three subsections review criticisms that target specifically the three conceptions of rationality section 1 presented.
Commitments Many take exception to the view that rationality requires egoistic maximization of selfinterest. They may concede that in trade, business, and markets rationality operates in a largely egoistic way. Still, they oppose applying the egoistic con end p.387 ception of rationality to all walks of life. They resist the “commodification” of education, health care, citizenship, and family life. Pace rational choice theory in the social sciences, they think that egoistic rationality should be constrained to the economic realm and not extended to social and political institutions such as the family and the law. Sen (1977 ) goes a step further. He argues against taking rationality to be the maximization of self-interest even in the economic realm. He observes that humans are
not motivated solely by self-interest, even in economic transactions where Edgeworth's assumption may seem to be a reasonable simplification. He observes that people are sometimes motivated by commitments to others—for example, to family, community, or class—and holds that such motivations are rational. Taking account of commitments, he claims, helps economists treat phenomena that fall squarely within their discipline's province, for example, the funding of public goods (such as roads, street lighting, and national defense) and the organization of a firm (including relations between labor, management, and shareholders). Duty may explain a citizen's honest revelation of a public good's benefit to him and may explain a worker's productivity in the absence of a comprehensive incentive system. Sen proposes revising economics' conception of rationality to allow for the commitments parents have to the welfare of their children, professionals have to the welfare of their clients, and citizens have to the welfare of their communities. Being motivated by commitment is compatible with a reasoned assessment of acts, and so is compatible with rational behavior in a broad sense (1977, 342–44).
Commitments to Teams and Plans If utility is defined in terms of choices and the standard of utility maximization is just a matter of making consistent choices, then that standard of rationality is open to the charge of being too weak. As Sen (1987 , 70) points out, an agent may consistently choose to frustrate his goals. Such behavior is irrational despite being consistent. Similar criticism arises if utility is defined in terms of preferences and the standard of having preferences that follow utilities is just a matter of having consistent preferences. An agent comparing risky options may have consistent preferences that are as if he assigns probabilities 40 percent and 60 percent to an event and its complement, respectively, when in fact his probability assignments are the reverse. Such preferences are irrational despite being consistent. In view of these powerful criticisms, this section examines utility maximization given a concept of utility that lets utility guide preferences and choices. It takes utility as a measure of strength of desire. Economists often have this concept of utility in mind when they take utility as satisfaction because they imagine that end p.388 satisfaction arises from attainment of any goal, including an altruistic one. Taking utility as degree of desire, utility maximization may still be regarded as a type of consistency between desires and choices, but it is a demanding type of consistency. Utility maximization so conceived faces some objections. Some theorists object that it makes sense only when desires are quantitative and yield comparisons of options. When desires are not quantitative and options are incomparable, it does not apply and supplementary principles are needed. See Chang 1997 . Imprecision and incomparability certainly limit application of the principle of utility maximization but do not challenge its credentials where it applies. So I move on to another, more biting objection.
Some theorists object that utility maximization, even with utility broadly conceived, ignores certain types of commitment. According to one version of the criticism, it ignores commitments to teams of agents. According to another version, it ignores commitments to plans. The first complaint is that maximization of personal utility, even if it accommodates altruistic goals, is too thin a conception of rationality to handle commitment to a team's goals. Consider, for example, a team of two altruists facing a coordination problem. Suppose they may coordinate in two ways, one of which is superior. Each altruist aims to participate in the coordination scheme the other follows. But what steers the pair to the superior coordination scheme if for some reason they cannot communicate? Each may identify with the team and pursue the team's goals and so achieve the superior form of coordination. According to Bacharach (forthcoming), such familiar behavior calls for standards of rationality going beyond maximization of personal utility. The second complaint is that utility maximization does not provide for the instrumental value of commitments to plans placing restraints on utility maximization. A plan involves a commitment to a series of acts, perhaps including cultivation of a character trait or a behavioral disposition. Some steps in a plan may not be utility maximizing but may be justified by the benefits of executing the whole plan. Utility maximization may be applied to long-term patterns of behavior rather than to every act as the occasion for it arises. Acts may be justified by being part of a utility maximizing plan rather than by being utility maximizing themselves. Restraint in maximizing utility at a moment may serve the goal of maximizing utility over a lifetime. Gauthier (1986 , chap. 6) advocates constrained maximization—roughly, acting on behavioral dispositions that are beneficial although they may call for nonmaximizing acts. Consider, for example, two people in a Prisoners' Dilemma, a decision problem in which cooperation by both benefits both but each has an incentive not to cooperate whatever the other agent does. Utility maximizing individuals fail to achieve cooperation. However, Gauthier points out that someone disposed to cooperate in a Prisoners' Dilemma profits from having the disposition if she is in a community of individuals who cooperate only with a person end p.389 having that disposition, even though her not cooperating still maximizes utility for her. According to his view, it is rational to cultivate the disposition to cooperate in those circ*mstances, and the cooperation to which the disposition leads is rational in virtue of the rationality of the disposition. McClennen (1990 , chap. 1) considers sequences of acts. He entertains cases in which a sequence maximizes utility although acts within the sequence do not maximize utility. In such a case his view is that the sequence's utility maximization makes rational the nonmaximizing acts in the sequence. A rational agent resolutely carries out the sequence of acts despite the incentives for deviation. For instance, suppose that you can win five dollars now if you turn down a dollar offered later. You resolutely form the plan to turn down the dollar offered later and are rewarded five dollars. Later you act on your resolution to decline the dollar offered although you already have been paid five dollars and would not lose that money if you were to accept the dollar. According to McClennen,
acting on the resolution is rational despite not maximizing utility. (See McClennen, chap. 12, this volume, on rules.) Bratman (1987 , chap. 2) draws attention to reasons that planning itself generates. Adopting a plan, he says, gives an agent a reason to follow the plan, a reason independent of utility maximization. If a driver is indifferent between two routes to his destination but settles on one, his having settled on that route generates a reason to take it. The reason does not depend on utility. Because of it, taking the other route is irrational although taking that route does not lower utility.
Bounded Agents The principle to satisfice rather than optimize draws attention to human limits and the costs of making decisions. Although suggestive and intriguing, it has not replaced principles of optimization in economics. It needs further articulation. An agent's aspiration level should be carefully regulated so that what counts as a satisfactory option adjusts appropriately to circ*mstances. In some cases of impending disaster, the level should drop dramatically to prompt a fast decision. In well-studied, routine cases where optimizing is easy, it should rise to the maximum attainable. Even after regulation of an agent's aspiration level, the principle to satisfice is still incomplete. Since the search for a satisfactory option determines which satisfactory option is found first, and so which option the principle recommends, the search procedure should be fully specified. Also, cases in which an agent finds multiple satisfactory options simultaneously require a supplementary selection principle. end p.390 Satisficing amounts to optimizing if an agent's aspiration level reaches an option at the top of her preference ranking, or if she is indifferent between satisfactory options. Some ways of articulating the principle to satisfice may make satisficing equivalent to a form of optimizing in all cases. Satisficing may be regarded as an optimization procedure whereby agents maximize utility but figure into utility calculations the cognitive costs of making decisions. Given limited cognitive powers, the costs of protracted deliberations are high, so shortcuts optimize all things considered. Raiffa (1968 , ix–x), for example, claims that his decision method is practical and takes account of human limits, although it aims at optimization. It registers the cost of deliberation when evaluating decisionmaking strategies and looks for an optimal strategy including that cost. Such optimization preempts the reasons for satisficing. Because it takes account of human limits, alternative principles of rationality are unnecessary. From the perspective of such theories, agents are bounded, not rationality, and satisficing, where defensible, agrees with rather than rivals optimizing. 2 Simon recognizes that satisficing may be formally equivalent to optimizing if decision costs are included in evaluations of options. He regards forms of optimizing that take human limits into account as forms of satisficing or forms of procedural rationality (1982 , 435). According to him, emphasizing procedural rationality rather than substantive
rationality nonetheless makes a practical difference to problem-solving strategies (1982 , 417–18). This defense of satisficing takes both satisficing and optimizing as decisionmaking procedures rather than as standards for evaluating decisions that an outside observer may apply after the decisions have been made. This distinction also arises in another defense of satisficing. Mongin (2000 , 95–104) argues against the attempt to reduce satisficing to optimizing. He claims that because of decision costs, optimizing is not always rational. Optimizing, he says, requires a decision to apply a principle of optimization, and optimizing in that second-order decision requires another, third-order decision to apply a principle of optimization, and so on. Thoroughgoing optimization therefore generates an infinite regress of decisions and therefore incurs prohibitively high decision costs. This objection arises, however, only if optimization is taken as a decision procedure rather than as a standard of evaluation. Meeting an optimizing standard of evaluation does not require a decision to apply an optimizing principle and so does not generate an infinite regress of decisions. Whether satisficing rivals optimizing thus depends on the interpretation of the principle to maximize utility. The principle may be advanced as a procedure for decision makers to follow, or as a standard for evaluating decisions. Since economic rationality traditionally addresses standards of evaluation for decisions rather than decision-making procedures, I take utility maximization as a standard of evaluation and put aside the issue of rational decision-making procedures. Under this interpretation, utility maximization more easily addresses the concerns of proponents of satisficing. Taking the principle as a standard of evaluation is a way of recognizing that cognitively limited agents are unable to make every decision using optimizing procedures. Slote (1989 , chap. 1) argues against optimizing and for satisficing on grounds independent of procedural advantages. He thinks that relentless optimization is immoderate. The virtue of moderation suggests satisficing, or being content with “good enough.” Slote holds that it is rational, out of moderation, to reject the optimal in favor of the good enough—that is, to decline knowingly an optimal option (1989 , 12). In such cases optimization is practical but satisficing is proposed nonetheless. Moderation is a virtue, but optimization may respond to its value. Consider an agent who appreciates the value of moderation. His utility assignment is affected by its value. If he must decide between two options and prefers one all things considered, including the value of moderation, then rationality requires its selection even if the other option is satisfactory. Moderation gives no reason for acting contrary to preference. Optimizing is not immoderate in this case. In general, arguments for departures from utility maximization tend to maintain that some acts are preferable to utility-maximizing acts. The arguments may be rebutted by letting utility encompass the good served by those acts so that they count as utility maximizing. In the case of principles of bounded rationality such as satisficing, recognition of cognitive limits, decision costs, and other factors motivating the principles suggest broader accounts of utility that include those factors. Those accounts tailor utility maximization to the abilities of cognitively limited agents. Similarly, broader accounts of utility may tailor utility maximization to the values of moderate agents. I conclude that the arguments for satisficing do not produce a genuine rival to the principle of utility maximization if the latter adopts a broad interpretation of utility.
3. Circ*mspect Maximization Some critics of economics' ideas about rationality promote alternative approaches to rationality. Gert (1998 , 39, 83–84) takes a deontic approach according to which an action is rational as long as it does not transgress commonsense principles such as those prohibiting gratuitous self-infliction of pain. Anderson (1993 , chap. 2) takes an action to be rational if it expresses rational values such as the value of health. Suppes (1984 , 187– 203) takes an action to be rational if it issues from reasoning. From a theoretical perspective, rather than make a fresh start, it is more end p.392 appealing to revise economics' claims about rationality to meet criticism and so extract if possible the truth they contain. It is more fruitful theoretically to build on their foundation than to start from scratch. Let other approaches' insights supplement rather than supplant economics' approach to rationality. I revise and defend utility maximization, with utilities for an agent taken broadly so that they register all her goals, including altruistic ones, and with utilities not defined in terms of preferences but rather introduced by utility theory in a way that makes them merely inferable from preferences and so still able to justify preferences. I take utility maximization as a standard of evaluation for decisions and advance it only for cases where options have utilities and so are comparable. The following subsections offer a brief defense of utility maximization. In Decision Space (2001) I more thoroughly examine it and methods of analyzing an option's utility.
Comprehensive Utility My way to meet criticisms of utility maximization is to let utility swallow up the reasons advanced for deviations. An option's utility responds to any reason to adopt the option. Any reason to adopt it is a reason to prefer it to other options and so to assign it a utility higher than theirs. Resistance to this defense of utility maximization argues that a broad interpretation of utility makes utility maximization trivial or even unfalsifiable. It may seem that utilities can always be assigned so that a person's choice maximizes utility. However, to make the interpretation of utility broad is not to license an arbitrary assignment of utilities. They must still represent a person's preferences. Putting aside operational definitions of preference and adhering to the ordinary, everyday concept of preference, a preference is not definitionally linked to a choice. It is possible for an agent to act contrary to a preference. This possibility ensures the possibility of an agent's failure to maximize utility. A principle that advises utility maximization is therefore not vacuous.
I apply my defense of utility maximization to criticisms appealing to commitments. If a person becomes committed to the welfare of others, then he gives acts that promote their welfare a higher utility rating. Commitment to others is not contrary to utility maximization. Comprehensive utility covers the value of honoring such commitments. Does maximization of utility thwart some community values even if utility is comprehensive? Utility maximizers take many steps toward cooperation and its benefits. In current interactions they take account of the consequences of their choices on the behavior of others in future interactions. If an agent's cooperation with others now encourages them to cooperate with him later, he has an incentive end p.393 to cooperate now. Also, societies, once established, may construct institutions to promote cooperation. Governmental authority may require cooperative measures that benefit all. Cooperation gathers momentum if a society's members have some measure of altruism that reduces incentives contrary to cooperation. Individuals with communitarian values may simultaneously maximize utility and promote social goals. Social dynamics, as Skyrms (1996 , chap. 3) explains, may account for the evolution of cooperation without novel standards of rationality. Does utility maximization nonetheless fall short of optimal forms of coordination in some cases where team members cannot communicate? There are two ways of taking this question. One asks whether adjustments in utility assignments may yield the optimal form of coordination. The answer is yes. Identifying with a team may make an agent's utility assignment conform with the team's. Then the agent's utility maximizing behavior yields the same result as novel principles of rationality concerning team action. Another way of taking the question asks whether, given the utility assignments agents typically have, utility maximization may yield coordination. The answer is again yes, given time for social dynamics to operate. Skyrms (forthcoming ) shows that rational people seeking an optimal form of coordination may succeed by learning to associate with those seeking the same form of coordination. Attending to association building, people may optimally solve coordination problems without reliance on novel rules of rationality. Does recognition of the value of planning require provision for deviations from momentto-moment utility maximization? Consider the following case in which an agent's lifetime utility apparently suffers because he maximizes utility moment to moment. Suppose that an agent is surrounded by people who pester him. He never retaliates when attacked, however, since such retaliation brings no benefit after he has suffered an attack. Those committed to the policy to retaliate if attacked despite the costs do better over a lifetime than he because of the deterrent effect of their policy. Although in this case nonmaximizing retaliation is rewarded, circ*mstances do not make such retaliation rational. If irrational retaliation is rewarded because people give berth to a person who retaliates irrationally, then a rational person cultivates that irrational behavior, as Pindyck and Rubinfeld (1989 , 479) observe. The principle to maximize utility acknowledges the rationality of cultivating that behavior without conceding the behavior's rationality. Rational agents cultivate dispositions and character traits that maximize lifetime utility. In environments where irrationality is rewarded, they may
cultivate dispositions and character traits that generate irrational behavior. This possibility does not demand revision of the principle to maximize utility. It remains the standard of rational behavior even in harsh environments. Finally, reconsider commitment to plans. The reasons for such commitments provide reasons to honor them. Sticking to a plan has benefits despite the attrac end p.394 tions of deviation. Failure to stick to a plan may thwart one's objectives. The constant search for improvements has cognitive costs. Agents with limited cognitive ability profit from resolutely following plans instead of calculating at each moment the act that maximizes utility. In cases where it is rational to follow one's plan, the cognitive costs of alternative strategies make following the plan utility maximizing. Recognition of cognitive limits brings utility maximization into alignment with rational planning procedures.
Recognition of Limits Let me return to the objection that utility maximization is too lofty a standard for humans. My way of meeting this objection is to acknowledge that nonmaximizing behavior may be rational unless an agent and her decision problem meet certain idealizations. According to the idealizations, the agent is cognitively perfect and circ*mstances for making a decision are perfect. The agent has been perfectly rational up to the time of the decision and will be perfectly rational subsequently, so that her decision need not compensate for irrationality elsewhere. Because economic theory uses rational behavior as an approximation to actual behavior, as Harsanyi (1977 , 16–19) and Sen (1987 , 68) observe, it may assume for simplification that rational humans maximize utility, although, strictly speaking, it advances utility maximization as a necessary condition of rationality only for ideal agents in ideal circ*mstances. Furthermore, despite the underlying idealizations, utility maximization may guide our decisions. It expresses a goal of rationality for them. Decisions that fall short of the goal are open to criticism even if nonideal conditions provide good excuses for falling short and so blunt a charge of irrationality. It may be a goal guiding decisions and decision preparations even though meeting the standard it expresses does not require adopting it as a decision procedure. A standard method of generalizing a theory that relies on idealizations is to rescind an idealization and then adjust the theory's principles for its absence. The section on satisficing, for example, discussed a way of revising the principle of utility maximization to accommodate cognitive limits. It suggested making an option's utility take account of decision costs. A way of doing this is to take an option as a possible decision. That is, if I am deciding whether to go to the store, one option is the decision to go to the store, not just the act of going to the store. The decision's utility includes the decision's costs—for example, the time and effort invested in forming the intention to go to the store. These costs lower the decision's utility and influence its place in a utility ranking of options.
To illustrate another generalization of utility maximization for nonideal cases, consider a decision problem in which options lack quantitative utilities because end p.395 desires are imprecise. As a result, the principle of utility maximization does not apply. Suppose, however, that despite not being quantitatively compared, the options are comparable. Then a generalization of utility maximization, formulated by I. J. Good (1952 , 114), applies. It involves what I call a quantization of beliefs and desires. This is just a probability and utility assignment that agrees with beliefs and desires. For example, if an agent wants a car more than a motorcycle, but the intensity of his preference is indeterminate, then one quantization may represent that preference as twice as intense as another does. Good's principle says that a rational decision maximizes utility under some quantization of beliefs and desires. According to it, utility maximization indirectly constrains rational choice even when beliefs and desires are imprecise. Although removing idealizations and revising the principle of utility maximization to accommodate greater realism is a challenging project, the two adjustments sketched show that progress is possible. In Realistic Decision Theory (Weirich forthcoming ) I pursue this project.
Generalization to Self-Support Besides generalizing the principle of utility maximization to handle cognitive limits, it is important also to generalize it to handle certain technical problems. Generalization brings out more clearly fundamental principles behind traditional forms of utility maximization. Suppose an agent confronts an infinite number of options and for each option another has greater utility. Then it is impossible to maximize utility. The principle of utility maximization needs generalization for such cases. Also, utility depends on subjective probability and thus information, and adopting an option itself supplies information. What is the appropriate body of information to use when computing an option's utility? It seems that an option's utility should be computed with respect to the information that it is adopted. But since different options then have utilities computed with respect to information differing about the option adopted, comparisons of options lack a common basis. Even if the number of options is finite, it may turn out that no option maximizes utility on the assumption that it is adopted. In terminology Jeffrey (1983 , sec. 1.7) introduces, it may turn out that no option is ratifiable. For example, suppose you are competing with another player in a game, and you have just two options. If your opponent is gifted at outwitting you, it may be that whatever option you adopt, its adoption is evidence that your opponent has foreseen and countered it so that the other option would have been better. The principle of utility maximization needs generalization for such cases, too. In Equilibrium and Rationality (1998, chap. 4) I propose a generalization that handles both problems. The first step defines an option's being self-supporting in a way that makes self-support subsume having maximum utility in standard cases but does not
require having maximum utility in general. The next step shows that in every decision problem, even the problem cases, some option is self-supporting. The final step argues that, under appropriate idealizations, rationality requires adopting a self-supporting option.
Conditional Rationality Refining the principle of utility maximization to meet objections also calls for a distinction concerning the scope of evaluation for a decision. We sometimes say that an agent's decision was rational given his beliefs but was irrational all things considered because his beliefs were irrational. Similarly, we sometimes say that an agent's decision was rational given the options he entertained but was irrational all things considered because he failed to entertain better options. The scope of a decision's evaluation may exclude some relevant factors or include all relevant factors. We may adopt more or less comprehensive standards of evaluation. Noncomprehensive standards may assess the rationality of a decision taking for granted some factor that comprehensive standards call into question. If idealizations do not remove mistakes an agent brings to a decision problem, utility maximization is a standard of conditional rationality only. It noncomprehensively assesses an agent's decision, taking for granted his utility assignment. The nonconditional standards that form the basis of a comprehensive evaluation must consider how a rational decision compensates for mistakes. In some cases departures from utility maximization may be appropriate. For example, a comprehensive evaluation may urge a decision that fails to maximize utility if the agent's utility assignment is irrational—say, if it springs from irrational anger. A maximizing decision may be rational taking for granted the agent's utility assignment but irrational from a comprehensive perspective that does not take for granted that condition of assessment. Working out the interplay of conditional and nonconditional standards of rationality is an exciting research project.
NOTES I thank Larry Alexander, Peter Vallentyne, and Xinghe Wang for helpful thoughts about my topic. 1. This distinction between procedural and substantive rationality differs from the end p.397 distinction made in ethics. For that distinction, see Hooker and Streumer, chap. 4, this volume. 2. Sen (1997 , 763, 768–69) distinguishes optimization from maximization and holds that maximization best assimilates satisficing. I do not distinguish optimization and maximization here. end p.398
chapter 21 LEGAL THEORY AND THE RATIONAL ACTOR Claire. Finkelstein
1. Introduction The application of rational actor theory to the law has been dominated by the law and economics movement. Law and economics began as a methodology for approaching legal questions; it suggested using the tools provided by economic theory to solve problems in manifestly economic fields like antitrust and taxation. But it quickly moved beyond methodology to acquire both a normative and a descriptive thesis of its own. The normative thesis is that the legal system should serve the goal of maximizing social welfare. The descriptive thesis is that the common law has developed according to this implicit economic logic, even if judges and other legal actors have not consciously tried to fashion legal rules with economic theory in mind. Law and economics thus regards law as a vehicle for maximizing utility, and it maintains that much of law can be explained as serving that function already. 1 In what follows, I shall leave law and economic's descriptive thesis about the nature of legal rules to one side and focus on its normative claim instead. On the normative side, law and economics shares a theory of value with act-utilitarianism in moral theory: Both posit some felicific notion like utility, well-being, happiness, or pleasure as the summum bonum for human beings. There are, however, several important differences between the two theories. First, legal economists have a end p.399 tendency to treat the notion of utility more generically than do ethical utilitarians. They include in it everything that is a source of value to an agent, whether or not it confers pleasure or happiness. Thus the legal economist might treat a person's altruistic feelings, aesthetic sentiments, and moral commitments as part of his utility, since these are all things a person values. Arguably, this makes the notion of utility vacuous, since it is not clear what could falsify the claim that a rational actor seeks to maximize his utility. (The economist, however, actually embraces this result, since he constructs an agent's utility function externally, namely from the choices he observes that agent making, rather than on the basis of the agent's internal psychological state.) For purposes of argument, then, I shall assume a somewhat narrower conception of utility, namely one that roughly corresponds to the idea of personal well-being. 2 A second difference between law and economics and act-utilitarianism has to do with the object of evaluation. While utilitarianism treats individual acts as the relevant units from which maximizing must take place, law and economics maximizes from the standpoint of legal rules. Law and economics thus proposes a two-tier theory: Legal rules should maximize social welfare, while individual acts falling under those rules need not be
themselves maximizing. In other words, law and economics is a kind of rule utilitarianism, where the relevant rules are legal, rather than moral, in nature. A third difference will provide the focus for our discussion. Law and economics is generally accompanied by a descriptive thesis about human nature, one that ethical utilitarians do not normally share. Legal economists tend to assume that the utilitarian theory of value to which they subscribe goes hand-in-hand with what we might call “rational actor psychology.” 3 According to this view, human beings are rational maximizers who reason instrumentally toward the attainment of their ends. One can understand why the legal economist thinks the utilitarian theory of value presupposes a rational actor model of human agency: If society's good consists in maximizing social welfare, it seems natural to suppose that individual good consists in maximizing personal welfare. But a moment's reflection should make clear that the psychological theory is neither presupposed nor entailed by the theory of value, since we need not think of the subjects of a utilitarian regime as capable of reasoning about their own good. We could, for example, ask what the best life for a cow would be and seek to maximize its utility by providing it with grassy fields and plenty of water. But we need not think cows capable of utility maximization on their own behalf, still less of anything like instrumental reasoning. The only requirement that social utility theory does impose on the creatures to whom it applies is that they be capable of experiencing pleasure and pain, since without this we could not meaningfully speak of their having any utility or well-being to maximize. Not only is rational actor psychology not inherently linked to the utilitarian theory of value, but the two are actually in some tension with one another. The end p.400 tension stems from the fact that when we maximize social welfare, we do not necessarily maximize the welfare of each individual member of society. In most cases, maximizing social welfare will result in some people faring worse than they otherwise would, even though other people fare better. This fact produces a quick route to the central claim I wish to make: If human beings are assumed to be rational maximizers, there is no reason to suppose they will be social maximizers as well. 4 And this suggests that a theory like law and economics that subscribes to a utilitarian theory of value at the same time that it adopts a rational actor psychology has some explaining to do. It must have a way of showing why the utilitarian theory of value it proposes should recommend itself to creatures with the particular constitution of rational maximizers. Indeed, I shall argue that rational actor psychology more naturally leads to a contractarian than to a utilitarian theory of legal rules. We must now back up, however, and take the long route to my conclusion. For the legal economist thinks he has a ready answer to the above concern. Unlike act-utilitarianism, law and economics does not actually require individual agents to maximize social utility. It is only legal rules that must maximize social utility, while individuals living under those rules need think only of their own good. Good legal rules will bring the incentives of individual agents into line with the requirements of social welfare. Thus individuals living under correctly crafted legal rules will maximize society's utility when they attempt to maximize their own utility. Utilitarianism and rational actor psychology can
co-exist in law and economics, then, because it is possible to adopt social rules that equip rational, self-interested agents with incentives to maximize social welfare. In what follows, I shall nevertheless take issue with this way of reconciling rational actor psychology and the utilitarian theory of value. I shall suggest that the tension between the two persists as a result of two central features of legal rules. First, I argue that legal rules must be created and administered by individuals who are themselves rational maximizers, and thus they must see themselves as personally advantaged by the rules they institute. Second, I argue that if human beings are rational maximizers, we ought to be able to justify a legal principle like utility maximization by showing that rules adopted in accordance with that principle would be in their interest. Yet there is reason to think that rational agents would not select rules that maximize social welfare. Indeed, as I shall argue, there is reason to suppose they would select rules premised on some sort of contractarian norm of agreement instead.
2. Rational Actor Psychology and Legal Actors The first problem with the legal economist's picture arises because rational actor psychology must be presumed to apply to what I shall call “legal actors,” namely people who create and implement legal rules. The tension between rational actor psychology and the utilitarian theory of value will thus resurface, since there is no reason to suppose that legal actors such as judges, jurors and legislators will maximize their utility by writing and implementing legal rules that maximize social utility. There would be no difficulty here if some other, perfectly utilitarian entity were to determine social and political arrangements for us. For in this case, like the cows about whose good we might reason, the motivations of the rule-makers need not mirror the motivations of the creatures to whom the rules would apply. If legal economists maintain that rational human beings would adopt and faithfully apply rules that maximize social utility, they should also be able to tell us why doing so should be in the interests of a rational legal actor. Some legal economists have noticed the difficulty, and they have sought to bring their account of the incentives of legal actors into line with their general account of individual motivation. Consider, in particular, the attempts legal economists have made to explain judicial behavior. Robert Cooter, for example, argues that public judges are primarily driven by the desire for prestige, and this leads them to act “in such a way that they would be chosen to decide cases by litigants and their lawyers if choice were allowed” (1983 , 107). They will thus maximize their personal utility if they faithfully apply the rules as they are written to the cases that come before them. And since those rules are ideally welfare-maximizing, judges will maximize social welfare when they maximize their own welfare. Richard Posner gives a slightly different explanation. He argues that the judicial utility function is based on the pleasure judges experience in voting on cases, since it is by rendering decisions that they have a sense of their own power. (1995, part 1, sec. 3). He analogizes this pleasure to the satisfaction members of the ordinary public receive from voting in elections—the sense of one's own importance that comes from the fact that one is entitled to decide. Consistent with this, he argues, is the fact that judges do not
particularly care about writing opinions. They see their function as residing in the decisions they render, rather than in the reasons they might articulate for their decisions. But given the job and salary security most judges enjoy, this does not explain why judges render decisions that adhere to legal rules, instead of deciding cases based on personal or idiosyncratic preferences. Why, that is, do judges care about the rule of law, when it does not maximize their personal welfare to do so? end p.402 Posner's answer is that adherence to the rules of judging is part of what makes the activity of judging pleasurable, much in the way that adherence to the rules of chess helps to make chess pleasurable (129). Judges adhere to the rule of law because they get more pleasure from following the rules, knowing it is this that defines their role, and hence their power, than they would if they decided cases for reasons of personal aggrandizement. He writes: The reason is not that judges have different utility functions from those of other people; it is that the utility they derive from judging would be reduced by more than they would gain from giving way to [various] temptations.It is the same reason why many people do not cheat at games even when they are sure they could get away with cheating. The pleasure of judging is bound up with compliance with certain self-limiting rules that define the “game” of judging.It is by doing such things that you know you are playing the judge role, not some other role; and judges for the most part are people who want to be— judges. (131) We could presumably tell a similar story about legislators. People who become legislators enjoy constructing rules that maximize social welfare. They are “publicspirited,” in the sense that their own utility increases to the extent they feel they are serving the public good. Indeed, carrying the parallel further, we might say that legislators derive their sense of their own power from their ability to create social legislation that incorporates their view of what the public good requires. For this reason, there need be no tension between the demands of individual welfare and those of social welfare. Presumably we could explain the behavior of jurors along similar lines. There are several problems, however, with this attempt to reconcile rule of law values with the hypothesis that legal actors are rational utility maximizers. I shall focus on Posner's account of judging, but similar criticisms could be made of the extension of that account to the other legal actors I have suggested. First, like the expansive approach Kaplow and Shavell take to the notion of utility, the rational actor psychology Posner's analysis assumes makes the claim that human beings are self-interested maximizers trivially true. If, for example, an agent has a strong moral compunction not to lie, cheat, or steal, this account will purport to explain the phenomenon by saying that he derives more utility from adhering to moral rules than from breaking them, since only then can he think of himself as playing a moral role from which he derives pleasure. Once again, this suggests that there is no motivational state that would falsify the hypothesis that legal actors are rational maximizers. Of course moral agents probably do derive pleasure from behaving morally. But we cannot use that pleasure as itself explanatory of the person's behavior if we wish to make a nontrivial assertion that the behavior is motivated by selfinterest.
Second, Posner's account maintains that adhering to legal rules has instrumental value for a rational judge. But no account of this form is likely to succeed, end p.403 for it fails to capture the experience of judges in relation to legal rules. Judges do not usually approach legal rules instrumentally. Instead, they regard themselves as bound by those rules, and do not feel free to ignore them if their personal utility would be enhanced by doing so. On Posner's story, in other words, we cannot explain the possibility of true internal commitment to the rule of law, since we cannot tell an instrumental story about such commitment. 5 Third, if Posner were right, judges would systematically violate legal rules when, in a given case, they would gain more utility from breaking the rules than from abiding by them. The widespread adherence to legal rules among judges, even in cases in which individuals would have much to gain from violating the rule, belies this suggestion. It is possible, of course, that judges receive such positive utility from following legal rules that it is nearly impossible for them to receive more from violating those rules. But, as I have suggested, this hypothesis is either trivial or highly implausible: It is trivial if we have no independent measure of a judge's utility than what we discover he happens to choose, and it is implausible if the notion of utility is taken seriously as an independent variable, and the claim is that following legal rules confers such great personal pleasure that no bribe or other extra-legal inducement could outweigh it. I thus return to my suggestion that insofar as legal rules and institutions must be created and implemented by rational agents, there is a tension between the utilitarian theory of value the legal economist espouses and the rational actor psychology to which he subscribes. For the theory to work, legal actors must be shown to be personally benefited by adherence to legal rules, and the few attempts that have been made to show this seem unpromising.
3. Rational Actors and the Social Contract The second problem with the legal economist's proposed reconciliation of rational actor psychology and the utilitarian theory of value requires us to look back to an earlier point at which general principles for the establishment of legal institutions would be selected. What we must ask is what principle or set of principles putative members of a legal regime would settle on as the basis for constructing the legal institutions under which they must live. That is, what kind of legal institution would individual maximizers see as most in their interest to adopt? Broadly speaking, there are three possibilities: They could adopt institutions that follow the principle of utility maximization; they could establish institutions based end p.404
on specific deontological commitments; or they could adopt legal institutions premised on the principle of mutual benefit. In other words, rational individuals might choose to make their legal institutions utilitarian, deontological, or contractarian. In what follows, I shall attempt to show why I think rational actors would not select the principle of utility as the theory of value for legal rules. I shall claim that they would also be unlikely to select deontological norms to structure legal rules. I conclude that rational agents selecting basic legal institutions are likely to select substantive norms of contractarian agreement. I shall attempt to demonstrate these claims by considering two examples of possible legal institutions. First consider a somewhat simplified version of an accident law regime proposed by Kaplow and Shavell (2002, III. C). Suppose rational actors are given the choice between two different regimes: allow people to drive and assign liability for resulting accidents according to fault, or allow people to drive but let losses lie where they fall. (A third option would be to disallow driving entirely, but for present purposes I will assume this option would not be adopted.) In the first regime, a victim could sue an injurer and recover damages for his losses if the injurer was negligent in causing the victim's injuries. In the second regime, no victim could sue any injurer. Which regime would rational actors choose? Suppose that each person could expect to be both an injurer and a victim exactly once in his life, so that the risks imposed between injurers and victims were “reciprocal.” In this case, Kaplow and Shavell argue, the parties would be indifferent between being able to sue for damages and leaving everyone to cover his own accident expenses. Now suppose there are high administrative costs involved in running a liability scheme, so that it is expensive for society to adjudicate and administer liability suits brought by accident victims. In this case, rational agents would choose to let losses lie where they fall, since the savings in administrative costs constitute a surplus that can be divided among the parties. The no-liability regime would leave at least one, and possibly all of the parties, better off than they would be under the liability regime—that is, the noliability regime would be pareto superior to the liability regime. Thus, they conclude, the claim that a tort liability scheme is a required part of corrective justice should be rejected. Rational agents would not choose to equip themselves with a fairness-based remedy when a different legal scheme would leave at least one person better off and no one worse off. Now consider the more realistic “non-reciprocal” case, the case in which the parties do not know when or whether they will be injurers or victims. In the absence of more specific knowledge of their situation, the parties should be indifferent between this case and the case in which the risk-taking is reciprocal. If there are high administrative costs to running a liability regime, Kaplow and Shavell once again argue, rational agents would choose the regime that leaves losses where they fall, since that regime would again be pareto superior to the end p.405 more expensive liability regime. As a general matter, Kaplow and Shavell assert that rational agents who do not know their ultimate positions under the regime they select will maximize their expected utility by selecting the regime with the highest social return. They would never choose a legal regime on grounds of “fairness.” Any such regime
could leave everyone worse off, since even the worst off group in an unfair regime might have a greater absolute share than they would in a regime where returns are distributed fairly. For this reason, Kaplow and Shavell may be correct that rational agents would not choose to equip themselves with a legal regime predicated on norms of fairness, and so to dispense with the suggestion that rational agents would select deontological norms as such to govern specific legal institutions. It hardly follows, however, that they would adopt the accident regime dictated by utilitarian norms. For the fact that a rational agent has a higher expected utility from a given legal regime does not necessarily mean that he would choose that regime, even if he is choosing strictly on the basis of self-interest. To see this, let us make use of an assumption that is common among legal economists, namely that a person who engages in risky behavior in some way benefits himself at the expense of others. The no-liability regime is thus one in which injurers improve their own position without having to pay for it, and victims are harmed without receiving compensation. Individuals living under such a regime are enrolled in a lottery, in which the injurers are the winners and the victims are the losers. By contrast, the liability regime would be one in which the parties' post-accident welfare would be restored to its preaccident state through court-ordered damage awards. In such a regime, there would be no winners and losers once compensation was paid. Now the question we might ask is: Would rational agents prefer the regime in which they gamble, hoping they will be nonliable injurers rather than uncompensated victims? Or would they prefer the regime in which they are assured of their pre-accident, baseline welfare, but in which they do not particularly stand to gain if they are not injured? In answering this question, let us assume that society would not allow people to drive without adopting one or the other of these liability regimes. That is, the “pre-agreement baseline” is the third option I mentioned above, namely that people do not drive at all. Arguably rational agents would prefer not to take a gamble that would potentially give them a lower payoff than they would receive from their ex ante baseline welfare. Indeed, their preference for protecting their baseline might be sufficiently strong that it would overcome even a higher expected utility from the regime involving a gamble. Rational agents, in other words, might be particularly disinclined to gamble with their baseline welfare, and be willing to forego the potential for even greater gains to protect that baseline. Why might this be the case? Consider how an agent contemplating an agreement for a future accident regime might reason. He might evaluate the rationality of the agreement by asking himself whether he would later be glad he entered that agreement or whether he would regret having done so. It is plausible to suppose that he would think himself well-served by the agreement if he is better off under the terms of the agreement than he would be had he never entered into the agreement in the first place. 6 That is, even if an agreement would require him to perform individual actions that leave him worse off, the agreement as a whole should leave him better off than he would be had he never entered into it. Let us call this condition the “benefit principle.” The requirement that an agent must receive a net benefit is a plausible threshold criterion for rational agents to apply in assessing options regarding the basic institutions of their society. The idea is that rational agents would not endorse basic institutions that they expect would leave them worse off than they would be in the absence of that institution. If rational agents did indeed apply such a test, they would not simply be concerned to
maximize their utility in selecting basic social institutions. Their interest in maximizing utility in this context would be subject to a constraint, namely that no matter how much expected utility an institution conferred ex ante, each agent should be reasonably confident that he would fare at least as well under that institution as he would in its absence. 7 The benefit principle should not be treated as a general test for the rationality of all agreements, plans or courses of action rational agents might adopt. For as a general condition of rationality, the principle would be much too strong: it would appear to rule out the rationality of homeowner's insurance, gambling (no matter how favorable the odds), and stock market investment. 8 But I am suggesting that such a strong condition would be rational with regard to agreements that govern the basic structure of society. Since rational contractors do not currently know the life choices they will be making, nor the kinds of preferences they are likely to have, they cannot count on ordinary calculations of expected utility to adequately protect their interests. Would the agreement to establish the no-liability regime satisfy the benefit principle in the non-reciprocal case? Under the no-liability regime, if a person happened not to become injured in an accident, he would regard himself as better off under this regime than he would be had he never agreed to that regime, since he would be able to enrich himself by engaging in risky behavior without paying the “price” for it. But he might turn out to be a victim without ever being an injurer. In such a case the no-liability regime would leave him worse off than he would be if he had never agreed to this regime, since his welfare would then be substantially below his pre-agreement baseline, and he would receive no compensation. Since the rational agent does not know which of these scenarios will occur, he cannot be sure that the agreement to establish the no-liability regime is one he will be glad to have entered. The no-liability regime thus appears to violate the benefit principle. What about the liability regime? No matter what position he happens to find himself in, an agent will regard himself as better off under the agreement to end p.407 establish the liability regime than if he had never entered into that agreement in the first place. For if he is an injurer, he will have to pay for the injuries he inflicts, but he will regard himself as better off for the opportunity to drive negligently, paying the price, than if he were not able to do so. (He presumably values the risky activity more than the discounted amount he must pay in compensation, otherwise he would not engage in the activity.) And if he is a victim, he is restored to his pre-accident baseline welfare, at the same time that he has the opportunity to engage in risky activities as a potential injurer. There is thus some reason to think the rational agent would prefer the liability regime. Indeed, he might prefer it sufficiently that he would be willing to accept it even if his expected utility were higher under the no-liability regime, as Kaplow and Shavell suggest. And if this is so, then we cannot assume that rational agents would systematically select the utility-maximizing solution. In saying that rational agents would select a legal regime that satisfies the benefit principle, I have not uniquely identified the legal regimes that rational agents would adopt. In theory, there might be many possible legal regimes that satisfy that principle. My claim is only that rational contracting agents would reject any legal regime that failed
the benefit test, and so we have reason to think they would not consistently choose the legal regime that maximized social welfare. In a fuller account, one would need to specify some further principle of selection that would allow the parties to choose from among the various eligible regimes. Presumably that principle would be consistent with contractarian assumptions. 9 My suggestion, then, is simply that rational agents would select the benefit principle as a minimum necessary condition for judging the acceptability of any agreement regarding legal regimes into which they might enter. Let us now apply the framework we developed in considering the accident law regime to a second question of legal policy, namely the death penalty. Assume that the death penalty has deterrent value over and above life in prison without parole. Would rational agents choose to include it in their punishment regime? If potential criminals are rational actors, the same level of deterrence can be accomplished either by combining a high probability of detection with low penalties, or by combining a lower probability of detection with more severe penalties. The legal economist therefore recommends that penalties be made as severe as possible, in order to minimize the use of costly resources in detecting and preventing crime. In general, the severity of punishment can be increased at less added cost than can the chances of detection. Thus economic analysis seems to point in the direction of using death as a penalty, if, as its proponents claim, the death penalty has greater deterrent value than life in prison without parole. But would rational agents setting up an institution of punishment select the form of that scheme by applying the principle of utility maximization? There are two reasons to think they might not. First, notice that the legal economist's argument places no upper boundary on the level of acceptable punishment. Sup end p.408 pose, for example, that torturing a person prior to executing him would increase the deterrent efficacy of the death penalty significantly. By the legal economist's lights, we should then opt for torture, since it would enable us to lower the amount spent on detection. It is probable, however, that the parties to an original social contract would treat certain kinds of very harsh penalties as off-limits, no matter what the utilitarian gains. The legal economist, by contrast, famously has difficulty explaining why certain morally impermissible punishments are unacceptable. Second, there is no reason to suppose that the person executed must be guilty of a crime in order for the death penalty to have deterrent efficacy. The legal economist's argument seems to sanction executing innocent agents. It is true that others will not be deterred from committing crimes unless they see at least a significant number of those who have committed those crimes receive a stiff enough sentence that they would rather abandon their plans than risk the associated punishment. But there is no need for the person punished actually to have committed the crime, strictly speaking. All that is really necessary is for the public to believe that he has done so. Now it might seem that the only way parties to a social contract would be able to eliminate such implications is if they were able to agree directly on a set of deontological principles that would rule out torture and sacrifice of the innocent. But it is unlikely that putative members of society could reach agreement on substantive moral principles in this way. Thus once again I suspect that Kaplow and Shavell are correct to think that
rational agents would not choose to govern their institutions according to specific deontological norms. The benefit of seeking to establish the basic structure by agreement among rational agents is to accommodate the fact that individuals have different views of the good and so cannot reach substantive agreement. How, then, would the parties to a social contract determine their principles of punishment? To determine whether rational agents would opt for the death penalty, let us once more apply the benefit principle, and ask whether a rational agent would regard himself as better off once he is subject to the death penalty, having lived with the benefits of a regime in which the higher deterrent effects of the death penalty are available, than he would have been in its absence. Does he prefer his life with the death penalty, now that he is subject to it, or the life he would have lived without that particular penalty? Presumably he would almost always prefer his life without the death penalty than with it, since he will not regard himself as better off once he is executed than he would have been had he never lived in a regime with the death penalty. That is, the agreement to institute the death penalty would have left him worse off, on balance, than he would have been in the absence of the agreement. The death penalty thus seems to violate the benefit principle and rational agents would therefore reject it. In this regard, the death penalty is like a kidney lottery society, in which each person has the option of entering into the following agreement: Should he suffer end p.409 dual kidney failure and require a kidney to survive, a member of the society will be chosen at random to supply a kidney in order that the victim of kidney failure might live. In such a case, the kidney lottery clearly passes the benefit test, since the person receiving the kidney is better off than he would have been in the absence of the lottery. But suppose he never does suffer kidney failure. Instead, someone else suffers kidney failure and our agent is chosen by lot to supply him with the needed kidney. The kidney lottery agreement would clearly fail the benefit test in that case, since the person chosen to supply the kidney will then wish he had never entered the kidney lottery. In this case, the lottery has turned out to be all cost and no gain. It is true that he might still benefit from the agreement, because should his one remaining kidney now fail, he will be the beneficiary of a lottery in which someone else must supply a kidney. But this is not a case of net benefit, since in the end he will be left with one kidney, which is what he would have had if he had never entered into the agreement in the first place. Arguably we should think of the death penalty as a kidney lottery. An agent who must be put to death by the state is like the person who suffers kidney failure under the lottery agreement: while he has undoubtedly benefited from the agreement to have the death penalty thus far, the agreement has not conferred a net benefit on him, since the fact that he will be put to death will now more than counter-balance any benefits he has received from the agreement. But how do we know this? Perhaps in the absence of the death penalty, everyone would die a violent death at someone else's hands at a very young age. So being able to live longer, and then be put to death, does not necessarily leave one worse off than one would have been in the absence of the death penalty. But in all likelihood, the availability of other severe punishments, such as life in prison without parole, would provide most of the deterrent benefits rational agents would require. The
marginal deterrent benefits death provides over other available punishments would almost certainly be insufficient to outweigh the loss of one's own life. One might object that the death penalty is distinctly unlike the kidney lottery, in that the decision to commit a crime is subject to choice, whereas suffering kidney failure is not. But once we allow that the state makes occasional mistakes in the administration of the death penalty, and that indeed it is quite impossible to run a death penalty without occasional loss of innocent life, the death penalty becomes precisely like the kidney lottery, since for an innocent person put to death, the loss of life is as much under his control as kidney failure. Moreover, even if we could ensure that no innocent people were put to death, there is reason to see even one's later decision to commit a crime as beyond one's control. This is because, from the ex ante perspective, a rational agent knows nothing of his future motiviations. The need or desire to commit a crime may be so compelling that his future self has no choice but to comply. Under either scenario, a rational agent is no more likely to enter a death penalty lottery than he would be to enter the aforementioned kidney lottery. end p.410 These examples are intended to suggest that utility maximization is not the only or even the most appropriate theory of value for a legal system designed by and for rational agents. Rational agents have reason to prefer legal institutions organized around the principle of mutual benefit, since each agent can be sure he will benefit in such a scheme. Thus if we take rational actor psychology seriously as a starting-point for legal theory, there is reason to question the claim that legal institutions would be organized around economic principles. It is at least as likely that they would be organized around contractarian principles instead.
4. Does Wealth Maximization Solve the Problem? In an article in the Journal of Legal Studies (1979 ), Richard Posner tries to distance law and economics from utilitarianism for precisely the reason discussed in the preceding section, namely that utilitarianism endorses the sacrifice of individual welfare for the sake of the greater good. He does this by suggesting that law and economics need not be premised on utility maximization after all. Instead, the legal economist suggests that legal institutions should seek to maximize wealth. Posner argues that wealth maximization better captures our intuitions about appropriate legal institutions. For example, a poor man who steals an expensive necklace he could not afford to buy may increase the total amount of utility in the world, since his wife, to whom he gives the necklace, may enjoy it more than the wealthy person from whom it was stolen. But the thief nevertheless decreases society's wealth, assuming he could not afford to purchase the necklace, since in an economic sense that means he does not value it. Thus arguably we can make sense of the law of theft in a wealth maximization system, but not in a system that seeks to maximize utility. It is worth revisiting the criticism I made of law and economics in
section 3, then, to consider whether rational agents would see their interests as adequately protected by a legal system that sought to maximize wealth rather than utility. There is some initial reason to think that wealth maximization would be an improvement in this regard. Posner argues that unlike utility maximization, wealth maximization requires voluntary exchange, since wealth cannot exist without markets. And voluntary exchange requires consent. As a result, law and economics will arguably no longer sanction the sacrifice of a smaller number of individuals for the sake of improving the well-being of others. For if society wishes to make use of the body or the possessions of one of its members, it will have to purchase the right to do so from him. As Posner puts the point: The great difference between utilitarian and economic moralityis that the utilitarian, despite his professed concern with social welfare, must logically ascribe value to all sorts of asocial behavior, such as envy and sadism, because these are common sources of personal satisfaction and hence of utility. In contrast, lawfully obtained wealth is created only by doing things for other people—offering them advantageous trades. The individual may be completely selfish but he cannot, in a well-regulated market economy, promote his self-interest without benefiting others as well as himself. (1979, 132) The thief, in other words, will have to purchase the necklace from its rightful owner rather than simply taking it, since a nonconsensual transfer would bypass a voluntary market. If Posner is correct, arguably a system predicated on wealth maximization would not permit institutions that left individuals worse off than they would be under their baseline welfare, since, as I have claimed, no rational agent would consent to such an institution. Thus it might be possible to acquire the benefits of a contractarian system within the framework of an approach that seeks to maximize social welfare, simply by measuring social welfare in terms of wealth instead of utility. One difficulty with this suggestion, however, is that economic analysis conceived in terms of wealth maximization is incomplete. The reason is that the rights respected in a market system are allocated prior to the operation of the market. That is, economic theory cannot itself specify the initial distribution of rights, since the theory actually presupposes such a distribution. In the case of the necklace, for example, we were assuming that the old owner of the necklace has rights to it that protected his entitlement when we said that the thief must purchase the necklace rather than just take it. But that system of rights has not itself been specified in economic terms; we have no apparent justification in terms of wealth maximization for allowing the old owner to reject attempts to remove the necklace from his possession without his consent. As Ronald Dworkin has convincingly argued, if the thief were willing to pay more for the necklace than the owner would accept to sell it, wealth would be maximized if a dictator were simply to give the thief the necklace, without requiring him to pay for it (1980 , 196–97). Posner argues against this point by suggesting that economic analysis is not indifferent to initial entitlements. There are two reasons for this. First, Posner argues that in general it makes sense to give each person the right to his or her own labor, since a person is likely to be more productive if he is his own master. In general, he thinks, there are economic reasons that bodies and labor should end p.412
be assigned to their “natural” owners, namely that natural owners are likely to value them most. Second, if there are transactions costs, an initial assignment of right to someone who values a good less than another person may be inefficient, since the parties may not be able to bargain for the exchange of the good. Thus Posner writes, “If transaction costs are positive, the wealth-maximization principle requires the initial vesting of rights in those who are likely to value them the most. This is the economic reason for giving a worker the right to sell his labor and a woman the right to determine her sexual partners. If assigned randomly to strangers these rights would generally (not invariably) be repurchased by the worker and woman respectively” (1979 , 125). The argument, however, is not a good one. Why assume that natural owners are likely to value their own labor or bodies more than others? Posner gives no defense of this claim, and on the face of it, it sounds more like an intuitive or moral idea than an economic one. The factory owner may indeed value the worker's labor more highly than the worker himself, in strictly economic terms, because he has the initial capital to make a large amount of money from the worker's labor, much more than the worker could ever be paid. Even a law firm will bill an associate's time at twice what they will actually pay the associate. The same might be said of a woman and her body: a pimp might make more from selling her body than she could ever make from selling her own. By substituting wealth maximization for utility maximization, Posner thus seems to save law and economics from the morally unacceptable results that come from its traditional association with act-utilitarianism. But it is not at all clear that he is able to do this without assuming a system of rights to property and to natural bodily entitlements. And if that is so, then law and economics can provide only a supplementary theory of how goods allocated according to non-economic principles can be fairly exchanged once initial entitlements are already in place. There is a second problem with the idea of wealth maximization, which Dworkin has also helpfully pointed out (1980 , 194–95). In order to be valuable, wealth must either be something of intrinsic worth—that is, an item worth having for its own sake—or it must be instrumentally valuable relative to something else that has value. Utilitarians think of utility (or happiness) as something worth having for its own sake. It is not valuable because it produces other things of value. Rather, all items that have value have it because they contribute to the total amount of utility. In this sense, the value of utility must be understood as self-justifying, since utility is the value in terms of which all other items of value are justified. If Posner is proposing that law and economics seek to maximize wealth instead of utility, he might be thinking of wealth in this same way: It is the ultimate value in terms of which all other items are valued. But this account cannot be defended, and Posner himself has admitted as much. In response to Dworkin's critique, he writes that wealth is a value because “it is conducive end p.413 to happiness, freedom, self-expression, and other uncontroversial goods” (1980 , 244). As Posner recognizes, if wealth is a value, it must be an instrumental value—something that is valuable because it allows one to acquire other items of value.
But if wealth is only instrumentally valuable, it is not at all clear why legal economists would want to maximize it. Is there any reason to think it would be better to maximize wealth than to try to maximize whatever is of intrinsic value directly? For example, assume in a given instance that we have reason to increase social wealth because we believe it will increase total societal happiness and not decrease any of the other goods we value. Is there any reason in such a case to seek to maximize wealth rather than happiness? Presumably not. And if not, why would we imagine that the benefits of wealth maximization change when maximizing wealth would maximize several other values as well? Moreover, there is a disadvantage to maximizing wealth as a proxy for maximizing other values: we do not know what quantity of these other values we will achieve if we just aim at wealth, whereas we can strike the balance we prefer if we aim at each value separately. Posner also suggests that by aiming at wealth rather than utility we can “thereby avoid[] certain well-known problems to which utilitarianism gives rise” (1980 , 245), presumably the moral problems we discussed earlier. But if we cannot aim at maximizing utility because it produces morally unacceptable results, then surely it does not improve matters to aim at wealth. For if maximizing wealth really does maximize utility, it will produce those same results, and if it fails to produce them, it can only be because we are not, in fact, maximizing utility. Not only does turning to wealth instead of utility fail to solve the second of the two problems with law and economics we considered above, 10 but it makes particularly clear just how serious the first difficulty is, namely the problem concerning the motivations of legal actors. 11 Is there any reason to suppose that an individual legal actor who is primarily concerned to maximize his own wealth would prefer rules that maximize social wealth? Not especially. It is possible that a rational legislator or judge would believe that his own personal wealth would be maximized if society were maximally wealthy, but the connection between the personal wealth of legal actors and society's wealth is not a particularly reliable one. A legal system that allowed legal actors to take bribes and kickbacks, for example, would probably maximize the personal wealth of legal actors, but it would be unlikely to maximize social wealth. Thus the gap between rational actor psychology and social welfare seems particularly significant if we substitute wealth for utility in the maximizing conception of value. end p.414
5. Conclusion I have argued that if human beings are rational, self-interested agents, they are most likely to favor institutions premised on the principle of mutual benefit, rather than on the maximization of social welfare. In this sense, rational actor psychology seems at least as likely to lead to contractarian as to utilitarian legal institutions. What it means for a legal institution to be based on contractarian principles will differ depending on the nature of the institution. But unlike a utilitarian approach, I have argued, the contractarian approach to legal institutions precludes legal rules that leave some worse off than they would be without those institutions. It also precludes the creation of institutions premised on the substantive moral commitments of any particular group that desires to force
compliance with their own views on others whose views diverge from theirs. What characterizes legal institutions premised on rational agreement is that the individuals that are subject to them can feel that those institutions have been established according to principles they accept as advantageous, even when the specific workings of those institutions are not. NOTES My thanks to Michael Abramowicz, Michael Green, Peter Huang, and Leo Katz for comments on earlier drafts and to Geoff Sayre-McCord for helpful conversations. 1. Although the term “welfare” is often taken to be broader than the notion of utility, I shall use the two terms interchangeably. 2. I do not think that this assumption in any way affects my disagreement with the legal economist. My criticisms would remain even on the broader conception of utility he proposes. 3. This psychological claim is not to be confused with what I referred to above as the legal economist's “descriptive” claim about the nature of legal rules. While it is a descriptive claim, it is a generic one about human motivation, one that is not particular to the legal economist. 4. Individual maximizers will produce a pareto optimal state of affairs. But that is not necessarily a state in which social welfare is maximized, assuming that each person's utility counts equally. 5. For a discussion of this point, in connection with Hart's account of the “internal point of view,” see Siegel 1999 , 1581 . 6. I have here adapted a criterion of David Gauthier's (1990 , 1994 ) for determining when entering into an agreement would be rational for an agent. But I make quite a different use of this criterion than Gauthier does, since I am applying it to test the rationality of enforceable agreements. There is arguably no need for such a test in this end p.415 context, because the agent can execute a plan he finds attractive on expected benefit grounds, without having to rely on his own willingness to perform a suboptimal action by way of compliance. Gauthier's test is a way for an agent to rationalize compliance without an agreement or plan that is not strictly speaking in his interest at the moment he must do his part. But arguably the criterion is a useful one for enforceable agreements as well. That is, we might think that the rational enforceable agreement is just the agreement the parties would have entered into in the absence of an enforcement mechanism. Enforcement, on this view, would be no more than a way of bringing somewhat irrational agents into line with the solution they would adopt if they were fully rational. I cannot, however, defend this account in any greater detail on this occasion. 7. Someone might argue that an individual utility function could be constructed whose maximization would satisfy the benefit principle, and that I have not therefore challenged the rationality of straightforward utility maximization at all. But the function being maximized would probably lack at least one of the features that economists typically impose on rational agency, namely transitivity, completeness, and continuity. 8. There are, however, ways to make the benefit more plausible as a general principle of rationality. As I have suggested elsewhere (Finkelstein 2003 ), it is plausible to think of
an ex ante chance of benefit as itself a benefit. And if this is so, then insurance and some gambles would be worthwhile: it would be rational to sign up for insurance, for example, if the ex ante chance of benefit were itself greater than the cost of the premium. Thus if the damage against which one is insuring is sufficiently great, and the chance that the damage will occur sufficiently high, then paying an insurance premium for an insurance policy that one does not end up needing nevertheless conveys a benefit. 9. One possible candidate would be David Gauthier's “minimax relative concession.” See Gauthier 1986 , Ch. V. Another would be Harsanyi's approach. See Harsanyi 1977 . 10. See supra section 3. 11. See supra section 2.
chapter 22 RATIONALITY AND EVOLUTION Peter. Danielson The most striking fact about the relationship between evolutionary game theory and economic game theory is that, at the most basic level, a theory built of hyper-rational actors and a theory built of possibly non-rational actors are in fundamental agreement. This fact has been widely noticed, and its importance can hardly be overestimated. Skyrms 2000 , 273 This chapter will focus on this “striking fact,” more generally: how game theory, the abstract theory of strategic interaction, supports new and powerful insights into the relation between rationality and evolution. First and most broadly, both rationality and evolution are optimizing processes working in a multiagent environment. I show in section 1 that the same modern methods that undermine naive evolutionary progressivism support the claim that rationality and evolution are isomorphic optimizing processes. This claim is striking, because, on their face, evolution and rationality are so different. While rationality tends toward a normative theory that fully applies only to extremely sophisticated, well-informed superagents, evolution manages to account for adaptation of the simplest organisms or even simple computer programs. The theoretical unification of these diverse phenomena is a major accomplishment for the relatively new field of game end p.417 theory. The second remarkable result of recent theorizing about these two processes is that they allow evolutionary and rational processes to be precisely compared. In section 2 I contrast a simple rational and evolutionary model and defend my focus on game theory, sketching the main concepts and variants of evolutionary game theory. The third interesting result is how the residual differences between evolution and rationality illuminate some crucial problems in the theory of rational choice: equilibrium selection and modular rationality. Setting various kinds of agents to play the same game under different information and selection regimes allows us to study precise differences. The research surveyed in this chapter comes from a broad range of disciplines, including mathematics, economics, biology, philosophy, political science, psychology, and computer science and engineering. This is remarkable in two respects. First, the
fruitfulness of the theories that unify evolution and rationality is indicated by their appeal to researchers in so many disciplines. Second, this unifying approach to social science has spread quite rapidly. For example, in 1967 the Encyclopedia of Philosophy (Edwards 1967 ) drew no connection between evolution and rationality, but by 1984, Robert Axelrod captures the convergence of game theory and biology by remarking that the Prisoners' Dilemma was the E. coli of social science. Later, “evolution” doesn't appear in the index of Elster 1986 's survey of leading contributions to rational choice theory, while half of Katz's recent collection on Evolutionary Origins of Morality (2000 ) concerns game theoretic arguments. While this chapter will not be historical, I will try to catch the interdisciplinary flavor of the material by quoting some of the original contributors of insights into the connection between rationality and evolutionary processes. In the fourth section, I consider the interdisciplinary aspect of the research that produced these results, with special attention to the role of models and simulations as a powerful research language; I sketch some two-level models that combine evolution and rationality. Fifth and finally, I return to the issue of normativity and consider whether the game theoretic approach to evolution is likely to encourage the evolution of rationality or something else.
1. Two Optimizing Processes Evolution and rationality differ in so many respects that it is remarkable that they could be thought of as nearly identical. As Jorgen Weibull (1998 , 3) points out, “A qualitative differencebetween evolutionary and rationalistic approaches is that while the second focuses on individuals and what goes on in their minds, the evolutionary approach usually instead analyses the population distribution of end p.418 behaviors (decision rules, strategies). One could say that the selection process replaces the mental process of choice made by rational players in classical non-cooperative game theory, while the mutation process replaces the mental process of exploring the strategy set and strategies['] payoff consequences.” Elliot Sober puts the general point forcefully: “Why bother to write about differences between two processes [deliberation and evolution] that are so obviously different?Deliberation involves a change in the composition of an individual; evolution effects a change in the composition of a population.Yet, in spite of these manifest differences, there seems to be an important isomorphism between the two processes[both are] optimizing processes.” This isomorphism relates evolution to rationality via what Sober calls “the heuristic of personification: If natural selection controls which of traits T, A 1 , A 2 ,, A n , evolves in a given population, then T will evolve, rather than the alternatives listed, if and only if a rational agent who wanted to maximize fitness would choose T over A 1 , A 2 ,, A n ” (1998, 408–9). Incidentally, this formulation of the analogy illuminates the (so-called) naturalistic fallacy, which Moore (1903 ) deployed to criticize some of Spencer's arguments (1884 )
from evolution to various values and which many have taken to block any normative significance for evolution (see Flew 1967 ). The generic fallacy is committed by arguments that attempt to output value content without a value input. However, we can see from the analogy above that neither rational nor evolutionary optimization need commit this error. Rational choice gets value out only by optimizing the value input captured by the agent's preferences. Conversely, what evolution optimizes is a value only if reproductive fitness is valued. The analogy becomes richer when we move to the perspective of multiple agents, organisms, or species. Darwin (1859 , 489) concludes The Origin of Species with this optimistic argument: “And as natural selection works solely by and for the good of each being, all corporeal and mental endowments will tend to progress towards perfection.” According to Sober's interpretation, Darwin is arguing that since selection maximizes individual fitness, it thereby maximizes group fitness. Sober points to the background assumption for this argument and indicates what goes wrong: The economists of Darwin's time tended to think that since a society is “nothing but” a collection of individuals, the society will maximize its well-being if each individual endeavors to maximize his welfare.Before one understands the structure of the tragedy of the commons, it perhaps sounds like a contradiction to be told that each individual, acting rationally and correctly foreseeing the consequences of his or her actions, will make a decision that leaves everyone worse off than he would be if some “irrational” choice had been made. 1 Perhaps it takes a rigorous decision theory on the one hand and the abstract picture of natural selection afforded by Fisher's theory on the other to bring the issues into focus. (Sober 1984 , 193) end p.419 “Selection may be relied upon to be a `progressive' force of evolutiononly when fitnesses are frequency-independent” (185). Hardin's tragedy of the commons model (1968) was designed to rebut straightforward optimization problem solving in strategic contexts. In Hardin's simple (but controversial; see Monbiot 1994 ) commons model, each of us rationally adds an additional grazing animal to a limited common field. However, the success of our strategies is frequency dependent. More grazing animals degrade the resource leading to poorer grazing for all. This is a many-player case of the problem seen most simply in the two-player Prisoners' Dilemma game. Each player's payoff depends on the number of players choosing that strategy; defecting alone pays the best, but when the other player defects as well, one's payoff drops to second worst. A classic real world tragedy of the commons is Gordon 1954 's model of the Canadian Atlantic cod fishery, which unfortunately has proven so accurate as a prediction that Atlantic cod is now practically extinct. Generalizing, the prevalence of environments with this interactive structure—now called social dilemmas—blocks the straightforward evolutionary route to perfection, or so it seems. In Dawkins's apt terminology, in these cases it is difficult to discern “God's Utility Function” (Dawkins 1996 , chap. 4). However there are some endowments so well suited for all environments, even these social dilemmas, that their evolutionary development should not be undercut by this argument. While grazers' and fishers' welfare will not be maximized by evolution in a
commons, their rationality should be. Rationality, after all, is the perfection of just those abilities useful for exploiting any situation, including social dilemmas. To use Lewontin's term (1978 ), as we expect populations subject to natural selection to “track” their environment, so should we expect agents so subject better to track their environment by means of improved cognitive equipment—that is, roughly, by means of rationality. Hence it is not surprising that modern progressive naturalists focus on the thin or formal epistemic and practical virtues associated with rationality, a doctrine that might usefully be labeled evolutionary rationalism. Strong forms of this doctrine are associated with Quine and Dennett—for example, these well-known epigrams: “Creatures inveterately wrong in their inductions have a pathetic but praise-worthy tendency to die before reproducing their kind” (Quine 1969 ) and “Natural selection guarantees that most of an organism's beliefs will be true, most of its strategies rational” (Dennett 1987 ). Ross and LaCasse (1995 , 488) state the central tenet more generally, linking it to the sense of rationality that will be the subject of this chapter: “An immensely complex conspiracy of pressures, some cognitive, some genetic, some that work through social and other environmental selection, drive systems that perpetuate themselves through control and feedback mechanisms to statistically converge towards rationality, in the sense captured by microeconomics.” Beniger 1986 makes a related argument historically, and changes the focus from organisms to organizations. Dennett's early statement (1978 , 16) of his po end p.420 sition makes the link to game theory explicit: “Game-theoretical predictions applied to human subjects achieve their accuracy in virtue of the evolutionary guarantee that man is well designed as a game player, a special case of rationality” (see Ross 2002 ). Evolutionary naturalism's general speculative thesis is strongly criticized in Stich 1990 and Kitcher 1985 . As we shall see in the next section, the convergence of evolutionary and economic game theory allows us to develop evolutionary rationalism in less speculative and more precise terms.
2. Isomorphism In this section I illustrate the isomorphic relation of evolution to rationality by means of a model that solves a pair of deep problems, one in evolutionary theory and apparently a different one for rational choice. I focus the discussion by limiting rationality to game theory and evolution to evolutionary game theory.
2.1.Sex and Fairness In most sexually reproducing species, the ratio between males and females is equal. This seems unproblematic—natural—to us, which simply reveals how naive our intuitions can
be. The biological problem is puzzling; generally females are the scarcity constraint on reproduction, so we should expect fewer males in an optimum sex mix. Darwin himself was deeply puzzled by this problem: “I formerly thought that when a tendency to produce the two sexes in equal numbers was advantageous to the species, it would follow from natural selection, but I now see that the whole problem is so intricate that it is safer to leave its solution for the future” (Darwin 1871 , quoted in Skyrms 1996 , 3). Skyrms (1996 , chap. 1) has us consider Darwin's puzzle in parallel with another: the elementary problem of dividing between two of us a simple good, such as a cake, when neither of us has any special claim. Here the problem is not what we ought to do— sharing equally seems obvious—but why we ought to do this. Skyrms argues that rational choice cannot answer this basic question about fairness. Game theory's central concept is the Nash equilibrium: “We have an equilibrium in informed rational self-interest if each of our claims are optimal given the other's claim. In other words, given my claim you could not do better by changing yours and given your claim I could do no better by changing mine” (5). But there is a problem. Each claiming half the cake is a Nash equilibrium, but so is me claiming 2/3 and you claiming 1/3. Skyrms explains, “There is a profusion of strict equilibrium solutions to our problem of dividing the cake, but we want to say that only one of them is just. Equilibrium in informed rational self-interestdoes not explain our conception of justice” (5). Skyrms constructs a model around the simplified game with three strategies for dividing a unit good shown in figure 22.1 . Notice that if the sum of the players' bids is greater than one, then each gets nothing, whereas if the sum is less than or equal to one, each gets their bid. In figure 22.1 the rows and columns are strategies and the payoffs (to the row strategy in the figure) represent reproductive fitness: the number of offspring expected. Let's think through the dynamics of this situation. If a player modestly demands 1/3, those who demand more will get more and will therefore reproduce more. Similarly, those who are greedy and demand 2/3 will lead to greater numbers of those who demand 1/3. Demand 1/2 “is a stable equilibrium. In a population in which everyone demands half the cake, any mutant who demanded anything different would get less than the population average. Demanding half of the cake is an evolutionarily stable strategy in the sense of Maynard Smith and Price, and an attracting dynamical equilibrium of the evolutionary replicator dynamics” (Skyrms 1996 , 11; we return to these concepts in sec. 2.2.1). Thus in the evolutionary model, equal division stands out among the several distributions. In this case evolution is able to solve the equilibrium selection problem “without imposing heroic cognitive requirements on players” (Bicchieri forthcoming ). The evolutionary model can produce a solution where rational choice fails because evolutionary payoffs are in reproductive fitness. In the simplest model of this dynamic process, the replicator dynamics, copies of each strategy are added to the population in proportion to its share of the total payoff. This means that the success of the strategies demand 1/3 and demand 2/3 will be limited by the negative feedback of encountering more copies of themselves. Now we can return to the problem with which we began, to see how evolution explains equal sex ratios. Supposing a parent's inherited tendency toward having more male or female offspring does not effect the number of children the
parent has, Fisher (1930 ) notes that the strategy “can nevertheless affect the expected number of grandchildren.In a population with a preponderance of males, a genetic tendency to produce more females would spread. There is an evolutionary feedback that tends to stabilize at equal proportions of males and females” (Skyrms 1996 , 7–8). The explanations based on this model are both stronger and weaker than they might first appear. Returning to the cake division problem, one might think the solution depends on the assumption of an initially equal population mix of strategies. But the argument is stronger, because the evolutionary model generalizes to other initial population mixes. This produces the diagram in figure 22.2 , where the vertices represent populations of pure strategies. (Since the populations must end p.422
Figure 22.1 Dividing the Fitness Cake add up to one, all possible combinations are restricted to this triangle, the so-called “simplex.”) The flows indicate that most population mixes of the three strategies lead to demand 1/2, except for the lower region, which gets trapped in a polymorphic mix of demand 1/3 and demand 2/3. This last possibility shows how the argument is weaker than one might expect. It involves no claim that equal division is a necessary outcome. Instead the equal outcomes form a basin of attraction the size of which is determined by the structure of the payoffs in the game. In the game under consideration, equality is likely but unequal outcomes are also possible.
Figure 22.2 Dynamics of Equality. Figure after Skyrms 1996 taken from www.ags.uci.edu/~jalex/latticegmodels. An interactive version of this model is available at www.ethics.ubc.ca/eame/eameweb/Skyrms/chapter 1.htm. Summing up, the evolutionary model solves both the biological problem about equal sex ratios and the problem concerning the counterintuitive results of rational choice in the divide-the-cake game. Equal sex ratios are an evolutionarily stable strategy in most populations, as are equal demands in the divide-the-cake game. The theoretical foundation of this result is the dynamic evolutionary model end p.423 that accounts for the two situations. I will not speculate whether this model accounts for our intuitions of fairness but will focus instead on the theoretical foundation in this chapter. Evolution is isomorphic to rationality only if we restrict the range of both concepts. This section will focus on practical rationality, or deliberation, and in particular I will focus on rationality in interaction. Here is where we would expect a rational and an evolutionary account to differ. Unlike theoretical rationality or practical rationality in a fixed environment, practical rationality where one's interests depend on the actions of others is the most complex problem of rationality. Elster (1984 , 18ff.) distinguishes strategic from parametric (or static) environments: “In the strategic or game-theoretic mode of interaction, each actor has to take account of the intentions of all other actors, including the fact that their intentions are based upon their expectations concerning his own. This was thought for a long time to involve an infinite regress, but frequently this is not the case. Using the concept of an equilibrium point it is possible to cut short the infinite regress and arrive at a uniquely definable and predictable course of action that will be chosen by rational men.” Therefore we assume the game theoretic account of strategic rationality, which is the subject of Bicchieri, chap. 10, this volume, and focus on the evolutionary variants of game theory.
2.2.Evolution as Evolutionary Game Theory
The ideas behind evolutionary game theory originate in Fisher 1930 , but the central technique was developed in Maynard Smith and Price 1973 , Maynard Smith and Parker 1976 , and Maynard Smith 1982 . Dawkins 1976 's brilliant exposition, radical formulations, and catchy phrases, like “selfish gene” and “genesmanship,” were very successful at propagating these ideas. Indeed, the second edition (Dawkins 1989 ) adds an appendix tracking the spread of some of these “memes” in the literature. Evolutionary game theory differs from economic game theory in several respects. Maynard Smith lists two: Paradoxically, it has turned out that game theory is more readily applied to biology than to the field of economic behavior for which it was originally designed. There are two reasons for this. First, the theory requires that the values of different outcomes (for example, financial rewards, the risks of death and the pleasures of a clear conscience) be measured on a single scale. In human application, this measure is provided by “utility”— a somewhat artificial and uncomfortable concept: in biology, Darwinian fitness provides a natural and genuinely one-dimensional scale.Secondly, and more importantly, in seeking the solution of a game, the concept of human rationality is replaced by that of evolutionary stability. The advantage here is that there are good theo end p.424 retical reasons to expect populations to evolve to stable states, whereas there are grounds for doubting whether human beings always behave rationally. (Maynard Smith 1982 , vii) Third, in an amusing contribution, Krugman (1996 , 2) claims, “The main difference between evolutionary theory and economics is that while economists routinely suppose that the agents in their models are very smart about finding the best strategy—an economist is always defensive about any model in which agents are assumed to act with less than perfect rationality—evolutionists have no qualms about assuming myopic behavior. Indeed myopia is of the essence of their view.” Evolution is one alternative to rationality, but we should avoid the temptation to make evolution the sole general purpose alternative to rational choice. In addition to evolutionary selection, agents may learn and be influenced by social norms. The application of evolutionary game theory to these latter cases is less well established. Historically, evolutionary game theory was introduced into biology, but the theory has been generalized on the one hand and limited to human agents on the other. I note three variants of evolutionary game theory.
2.2.1. Biological Evolutionary Game Theory The central notion in biological evolutionary game theory is an evolutionarily stable strategy (ESS) introduced by Maynard Smith and Price attempting to explain why most intraspecific animal conflicts are “limited war” rather than “total war” types: “The concept of an ESS is fundamental to our argument; it has been derived in part from the theory of games, and in part from the work of MacArthur and of Hamilton on the evolution of the sex ratio. Roughly, an ESS is a strategy such that, if most of the members
of a population adopt it, there is no `mutant' strategy that would give higher reproductive fitness” (1973, 15). Formalizing this definition links the ESS with game theory's Nash equilibrium concept (see Bicchieri, chap. 10, this volume). Strategy x is an ESS if, for every strategy y distinct from x and utility function u (where u(x,y) is the utility of “playing” x against y), 1. u(x,x) ;eg u(y,x) 2. If u(x,x) = u(y,x) then u(x,y) > u(y,y) According to Samuelson (1997 , 40), “This, the original, definition of an ESS offered by Maynard Smithhas the advantage of making it clear that the ESS condition is the combination of a Nash equilibrium requirement, given by (1) and a stability requirement, given by (2). The latter ensures that the ESS [x] can repel mutants.” In addition to an account of equilibrium selection, evolutionary game theory “provides an explanation of the dynamics of the selection process, something end p.425 which [game theory's] refinement program cannot do” (Bicchieri forthcoming ). Returning to the simplest case, the replicator dynamics, “every ESS is a strongly dynamically stable (or attracting) equilibrium in the replicator dynamics” (Skyrms 2000 , 274). As Samuelson notes, “The best-known form of the replicator dynamics was introduced by Taylor and Jonker 1978 and by Zeeman 1992 but similar ideas have been used in earlier applications. Hofbauer and Sigmund [1988 ] discuss one of the earliest analysesin which the mathematician V. Volterra constructed a dynamic model to explain population fluctuations of predator and prey fish species in the Adriatic Sea” (1997, 18). Strictly speaking, the replicator dynamics models a simplified ecological approach to dynamics, as the types of interactor in the population are fixed. Axelrod (1981 ) distinguishes an evolutionary approach where new types of interactor can enter the population. He later pioneered using the genetic algorithm to implement this more complex model in Axelrod 1987 . (Axelrod's work is discussed further in sec. 4.1 below.) The genetic algorithm, pioneered by Holland (1975 ), mimics the mechanism of biological evolution to generate new variants as well as to test them by differential selection. Typically pairs of successful functions, composed of vectors of bits, are combined by crossing over part of each parent vector; vectors are also altered by point mutation. The genetic algorithm allows a model to explore a much larger space of agent interactions than is possible by enumeration of agent types combined with the replicator dynamics.
2.2.2. Economic Evolutionary Game Theory
Where biological evolutionary game theory is intentionally broad in the scope of its agents, economic evolutionary game theory focuses more narrowly on explaining human action. Economic evolutionary game theory can be seen as a reply to Maynard Smith's point quoted above about the replicator dynamics applying better to nonhuman organisms. Mailath (1992 , 268) agrees: “There is nothing in economics to justify the replicator dynamic.” Therefore, economic evolutionary game theory modeling is based on a human learning dynamic. The “model consists of a large population of myopic and unsophisticatedplayers.[An unsophisticated player] does not conduct his calculations under the assumption that the other players are similarly reoptimizing.[The players] can observe the history of plays” (Mailath 1992 , 261). Samuelson 1997 makes a similar distinction between economic and biological evolutionary game theory.
2.3. Evolutionary Generalism At the other extreme, Skyrms is an “evolutionary generalist” (D'Arms, Batterman, and Gorny 1998 ) who aims to model evolutionary processes in the most abstract form. He explicitly states that his models are sufficiently general to cover both biological and cultural evolution: “[Fair division's] strong stability properties guarantee that it is an attracting equilibrium in the replicator dynamics, but also make the details of the dynamics unimportant. Fair division will be stable in any dynamics with a tendency to increase the proportion (or probability) of strategies with greater payoffs.For this reason, the Darwinian story can be transposed into the context of cultural evolution, in which imitation and learning may play an important role in the dynamics” (Skyrms 1996 , 11). Similarly, Dawkins 1983 and Dennett 1978 , chap. 5, make claims for the universality of Darwinism. Note that when we generalize evolutionary models to include social as well as biological evolution, we can no longer always rely on Maynard Smith's objective Darwinian alternative to subjective utility functions. This is because in social evolution reproductive success can be mediated by agents who choose which agents' strategies to imitate. Indeed, since these copiers need to be able to compare different agents' utilities, they require a more demanding subjective utility function (Bicchieri forthcoming ).
2.4. Why Isomorphism Is Important To return to our opening quotation, Skyrms draws an important conclusion from the way game theory shows evolution and rationality to be isomorphic: The most striking fact about the relationship between evolutionary game theory and economic game theory is that, at the most basic level, a theory built of hyper-rational actors and a theory built of possibly non-rational actors are in fundamental agreement. This fact has been widely noticed, and its importance can hardly be overestimated. Criticism of game theory based on the failure of rationality assumptions must be reconsidered from the viewpoint of adaptive processes. There are many roads to the Nash
equilibrium concept, only one of which is based on highly idealized rationality assumptions. (Skyrms 2000 , 273)
3. Differences Having established the strong similarity of evolution and rationality, their differences become interesting. Elster 1989 , chap. 8; Skyrms 1996 ; Skyrms 2000 ; and Sober 1998 all survey differences between rationality and evolution. To return end p.427 again to our opening quote, Skyrms continues: “However,the situation ismore complicated than it might at first appear. There are aspects of accord between evolutionary game theory and rational game theory as well as areas of difference. This is as true for social evolution as for biological evolution. The phenomena in question thus have considerable interest for social and political philosophy.” (Skyrms 2000 , 273). “Throughout a range of problems associated with the social contract, the shift from the perspective of rational choice theory to that of evolutionary dynamics makes a radical difference. In many cases, anomalies are explained and supposed paradoxes disappear” (Skyrms 1996 , xi). We will consider three differences.
3.1. Symmetry “A single population evolutionary setting imposes a symmetry requirement which selects Nash equilibria which appear implausible in other settings” (Skyrms 2000 , 273). Skyrms (1996 , x) “shows how evolution imposes a `Darwinian veil of ignorance' that often (but not always) leads to selection of fair division in a simple bargaining game. In contrast, rational decision theory leads to an infinite number of equilibria.” This case was discussed in section 2.1 above.
3.2. Dominated Strategies “Refinements of the Nash equilibrium are handled differently. Standard evolutionary dynamics, the replicator dynamics, does not guarantee elimination of weakly dominated strategies. In a closely related phenomenon, when we consider extensive form games, evolutionary dynamics need not eliminate strategies which fail the test of sequential rationality” (Skyrms 2000 , 273). In a chapter on “Commitment,” Skyrms (1996 ) uses the asymmetrical game of Take It or Leave It to demonstrate these differences. This game puts player 1 in a privileged position, able to make player 2 a fair (demand 5) or an unfair (demand 9) offer, which the latter must then accept or reject (getting nothing). In this
sequential ultimatum game, (see figure 22.3 ), human experimental subjects moving first often make fair initial offers, while the rational strategy is to offer the second player the smallest possible share—the solid lines, leading to the {9,1} payoff. Various explanations of these results do not go very deep. If humans prefer fair outcomes, why do they have these preferences? “Appeal to norms of fairnesshardly constitutes an explanation in itself.Why have norms of fairness not been eliminated by the process of evolution?” (Skyrms 1996 , 28). Skyrms constructs an evolutionary model of the ultimatum game. This model end p.428
Figure 22.3 Ultimate Game. is more complex than that discussed in section 2.1, since each player needs, when playing the role of player 2, to be able to react to the two strategies open to player 1.Thus there are eight possible strategies, including Gamesman, who demands 9 and accepts all offers, Fairman, who demands 5 and accepts 5 but rejects 9, and the counterintuitive Mad Dog, who demands 9 but rejects 9 and accepts 5. The first two are the strategies that interest us initially, as the first is the recommendation of rational choice theory, and the second, although it would explain the experimental results, is rationally defective. By rejecting demands of 9 (offers of 1), Fairman threatens to choose what is worse (0 over 1) and thus fails the test of “modular rationality,” which requires that a strategy “specify a rational choice at each choice point” (Skyrms 1996 , 24). Skyrms shows that although the Fairman strategy is dominated by Gamesman—the latter never does worse and sometimes better—some initial populations will allow Fairman to persist (see figure 22.4 ). This is a case where misleading results can follow from oversimplified evolutionary models. Skyrms points out that, because it fails to generate the strategies that would do better, the replicator dynamics can leave dominated strategies in place. In this way, the replicator dynamics are conservative. He shows that introducing crossover (with a genetic algorithm; see above, sec. 2.2.1) need not correct this bias; mutation is needed to test robustness fully. Thus he stresses the important differences, with respect to contrasting evolution and rational choice, between variations of the genetic algorithm that are often treated as merely technical variants. Nonetheless, even
the more refined evolutionary model will allow dominated strategies to persist. This sounds like good news for Gauthier (1986 ) and McClennen (1990 ), who have defended dominated strategies of commitment as the doctrines of constrained maximization and resolute choice, respectively. Perhaps evolutionary game theory can add new life to these theories, which have not been well received as contributions to economic game theory; see, for example, Binmore 1994 . We look at one way this might play out in the concluding section. end p.429
Figure 22.4 Dynamics of Ultimatum Game. Figure after Skyrms 1996 taken from www.ags.uci.edu/~jalex/latticegmodels. An interactive version of this model is available at www.ethics.ubc.ca/eame/eameweb/Skyrms/chapter 1.htm. Samuelson (1997 , 34) suggests that these results will have general impact on the theory of rational choice: “The failure to eliminate weakly dominated strategies is characteristic of a wide variety of evolutionary models. I suspect that if there is any general implication to emerge from evolutionary game theory, it will be that we should be much less eager to discard weakly dominated strategies and should place less emphasis on dominance arguments in selecting equilibria. This in turn may lead to reconsideration of applied work in a wide variety of areas.”
3.3. Actual Fitness and Intended Outcomes A third difference between rationality and evolution has to do with the payoffs. “Rational choice is concerned with the intended outcomes of action. Selection mechanisms operate through actual outcomes. In explanations of animal behavior, where intentions have at best a minimal place, actual outcomes must bear most of the explanatory burden. It is more controversial which mechanism is the most important in the study of human action” (Elster 1989 , 71). Binmore (1998a , 33) appeals to this difference to explain the previous result: “The evolutionary argument holds together `tolerably well' when used to defend the notion of a
Nash equilibrium but the abstract form of the argument comes under strain when applied to subgame-perfect equilibria because this notion depends on postulating the players would behave consistently even under circ*mstances that never actually occur. The orthodox neoclassical notion of consistency then parts company with the evolutionary story since the irrational plans can only be selected against if they are sometimes brought into play.” This third difference lets us correct what may be a misunderstanding of the impact of evolutionary game theory literature. Starting with Axelrod (1984 ), there has been a tendency of some to see evolutionary models as generally supporting end p.430 cooperation in cases—like the Prisoners' Dilemma—where rationality would not. However, as Sober (1998 , sec. 4) points out, the third theoretical difference allows that evolutionary selection will sometimes fail to cooperate where rational agents, looking ahead to the consequences of their own defection, would cooperate. Generalizing this point, I have argued elsewhere (Danielson 1998a ) that one should not identify evolutionary models with assumptions about positive correlation of similar agents, which can be used to derive more cooperative outcomes.
4. Simulations and Multilevel Models Roughly, the contrast between rationality and evolution projects onto methods, with evolution characterized more by simulations and rationality by more formal models. We begin with Axelrod's influential work and the criticism it has attracted, move on to general skepticism about evolutionary simulations, and end more constructively with some recent work on multilevel simulations. Incidentally, as Dawkins (1976 , chap. 4) points out, the use of simulations is related to our topic in a more general way. For the same reasons of efficiency that we researchers build computer simulations to predict what might happen under various assumptions, so too evolution is drawn to produce agents that model their environments—that is, to become rational, responding to indirect reinforcements, and in Popper's famous phrase, “let [their] hypotheses die in [their] stead” (quoted in Dennett 1995 , 375).
4.1. Axelrod While evolutionary biology provided one path into contemporary evolutionary game theory, the political scientist Robert Axelrod's tournament experiments and simulations (1981 , 1984 , 1987 , 1997 ) provided another. Axelrod invited the research community to submit strategies (programmed in FORTRAN) for an indefinitely repeated Prisoners' Dilemma, which he played in a round-robin tournament. He then applied the replicator dynamics to simulate further rounds of the tournament. Critics of Axelrod lament two
aspects of his influential work: first, that he simulates what should be obvious from theory—that is, that there are infinitely many equilibria in the repeated Prisoners' Dilemma (the so called folk theorem). Second, that he overemphasizes one equilibrium strategy: tit-for-tat. The second is an important criticism, but the first is unfair. Axelrod introduced important new methods and initiated an interdisciplinary dialogue based on them. Notice, for example, that Axelrod (1987 ) initiated the use of programmed strategies, which was influential in computer science, as well as the use of genetic algorithms to generate strategies, which brought a previously disconnected field into closer contact with economics and biology. Unfair, because some of the best results refining Axelrod's work do not derive from the pure theory that he “should have” deployed, but by means of the new methods that he pioneered. For example, Binmore and Samuelson (1992 ) build populations of finite state machines. Binmore (1998b ) assesses Axelrod's contribution: The folk theorem of game theory proved by several authors simultaneously in the early fiftiesdescribes in precise detail all of the outcomes of a repeated game that can be sustained as equilibria. However, although Axelrod didn't discover the folk theorem, I believe that he did make an important contribution to game theory, which has nothing to do with the particular strategy TIT-FOR-TAT nor with the mechanisms that sustain any of the other equilibria in the indefinitely repeated Prisoners' Dilemma. He did us the service of focusing our attention on the importance of evolution in selecting an equilibrium from the infinitude of possibilities whose existence is demonstrated by the folk theorem. Other game theorists may protest at my recognizing someone who knew no game theory at the time he made his contribution and still resolutely ignores gametheoretic commentary on his work, but it is inescapable that the evolutionary ideas pioneered by Axelrod now provide the standard approach to the equilibrium selection problem in game theory.
4.2. A Methodology: Multilevel Modeling So far we have emphasized the similarities and differences between evolution and rationality. Stepping back, it is obvious that the complete picture must include both; evolution has historically produced some rational agents. So the question arises, how might one model the interaction between evolution and rationality? Given the difference in time scales of the two processes, a natural way to approach the relation of rationality and evolution is by a two-level model. Binmore (1994 , 151) suggests a framework: The principal agent problem involves a principal with certain aims who can only act through agents who may have aims of their own. The principal therefore seeks to design an incentive scheme that minimizes the distortions resulting from having to delegate to the agents. Nature, as principal, is blind, but her agent, hom*o economicus, can see his environment. Thus, although Nature loses fine control being unable to modify his environment directly, she gains access to information that would not otherwise have been available to her. If the environment changes sufficiently rapidly, the gains will outweigh the losses. end p.432
This framework suggests using an evolutionary mechanism to generate and select agents' preferences while a rationality mechanism uses those preferences to select strategies. Levine (1998 ), Sethi and Somanathan (2001 ), and Danielson (2002 ) explore questions of reciprocity and fairness employing models of this form; Harms (1997 ) applies a twolevel model to questions in evolutionary epistemology. Of course, simulation is not limited to two levels of modeling, although the more complex the simulation the more difficult it is to understand its output. Epstein and Axtell 1996 is an innovative exploration of a multilevel simulation, incorporating not only evolution of agents who make (boundedly) rational choices (of movements) but also evolution of their immune systems and social identity tags.
4.3. Computational Differences We have focused on comparing rationality and evolution rather narrowly as formal accounts of practical rationality in social interaction, abstracting away most computational problems by using equilibrium concepts. But if we broaden our contrast, rationality looks weaker. In their provocatively titled “Better than Rational: Evolutionary Psychology and the Invisible Hand,” Cosmides and Tooby (1994 ) argue, One point is particularly important for economists to appreciate: it can be demonstrated that “rational” decision-making methodsare computationally very weak: incapable of solving the natural adaptive problems our ancestors had to solve reliably in order to reproduce.This poor performance on most natural problems is the primary reason why problem-solving specializations were favored by natural selection over general purpose problem-solvers. Despite widespread claims to the contrary, the human mind is not worse than rational (e.g. because of processing constraints)—but may often be better than rational. On evolutionarily recurrent computational tasks, such as object recognition, grammar acquisition, or speech comprehension the human mind greatly outperforms the best artificial problem-solving systems that decades of research have produced, and it solves large classes of problems that even now no human-engineered system can solve at all. However, we should not draw too practical a conclusion from this claim. It doesn't follow that one can better engineer a general purpose robot or a chess-playing automaton by applying evolutionary rather than rational techniques. The competition between these methodologies in artificial intelligence is a matter of lively controversy, ranging from strong claims for evolutionary techniques (Koza 1992 ; Pfeifer and Scheier 1999 ) to this dry “logicist” reply in a leading AI textbook: “Like neural networks, genetic algorithms are easy to apply to a wide range of end p.433 problems. The results can be very good on some problems, and rather poor on others.Denker's remark `neural networks are the second best way of doing just about
anything' has been extended with `and genetic algorithms are the third'” (Russell and Norvig 1995 , 621).
5. Normativity Although we began with a discussion of the normative limitations of appeals to evolution, so far we have considered evolution and rationality as mechanisms without regard to their normative or descriptive use. I turn to this issue in this section and close with some intriguing conjectures about the effect of theory on the evolution of human rationality.
5.1. Normative Rationality versus Descriptive Evolution As other chapters in this volume discuss the normative status of rationality, here we focus on how having an alternative evolutionary mechanism influences this issue. We have already noted in section 1 that the isomorphism between evolution and rationality affects their role in normative arguments. Rational choice only transfers value to alternatives valued by preferences; evolution optimization matters only if one values fitness. Critics of normative appeals to evolution were typically aiming at more substantial theses, such as the claim that complexity is good because organic complexity is the goal of evolution. The standard view is that while rational choice is normative, evolution is merely descriptive. Nonetheless, the connections between rationality and evolution do give the latter a new normative significance. First, there is the philosophical question why rationality but not evolution should be seen as normative, given their strong formal similarity. One can, of course, criticize evolutionary mechanisms for not selecting a global optimum when faced with a local optimum. But conversely, one can criticize the rationality mechanism for limiting an agent to the alternatives she conceives of. Similarly, the driver of evolutionary models is reproductive fitness, the value of which is open to the antinaturalistic challenge “Why should I care about being copied?” (Or more generally, what de Sousa (1980 ) calls Oscar Wilde's Boomerang: “The first duty in life is to be as artificial as possible.”) But the analogous rejoinder end p.434 here, leading to “Why should I care about my own preferences?”, is only disallowed because preferences are defined as determining behavior. If we consider a more complex case, where an agent finds herself choosing to smoke or drink against her acknowledged interest in health or social life, matters are less clear (Varner 1998 , Rachlin 2000 ). Second, recent normative theory gives rationality a greater normative role and this gives evolution a new role as well. Starting with Rawls's (1971 ) influential contractarian scheme to make the theory of justice part of the theory of rational choice, Gauthier (1986 ) and Binmore (1994 , 1998a ) develop and refine the role of rationality in moral theory.
Therefore, when evolution modifies the theory of rational choice, it has normative consequences as well. For example, Skyrms (1996 , 6) argues that the indeterminacies of rational choice also infect normative theories built upon it. Following his argument that straight rational choice fails to find a solution to the simple bargaining problem (discussed in sec. 2.1), he claims that the social contract theories of Rawls 1971 and Harsanyi 1955 , which invoke the device of the veil of ignorance, behind which neither of us knows which piece of cake he will get, also fail to account for the salience of the equal outcome. Third, as we have seen, evolutionary arguments may change our conception of rationality. Samuelson (1997 , 3) notes the philosophical uncertainty of key rationality concepts: “For some, game theory is a normative exercise investigating how decisions should be made.The intricacy of this problem grows out of the self-reference that is built into the notion of rationality, with all of its attendant logical paradoxes. The richness of the problem grows out of our having no a priori notion of what it means to be rational, substituting instead a mixture of intuition, analogy and ideology. The resulting models have a solid claim to being philosophy and art as well as science.” Evolutionary equilibrium selection may influence these intuitions. For example, “the restrictions to be placed on out-of-equilibrium behavior may appear to be obvious in the Chain-Store Game, but more formidable challenges for an equilibrium refinement are easily constructed, involvingmore intricate challenges to our intuition as to what is a `good' equilibrium. [This literature] has produced an ever-growing collection of contending refinements, many of which are quite sensitive to various fine details in the constructions of the model, butvery little basis for interpreting these refinements or choosing between them” (Samuelson 1997 , 9). Fourth, while evolutionary game theorists have been content to see the theory as “a positive exercise in investigating how decisions are made” (Samuelson 1997 , 4), we should note the interdependence of positive and normative considerations. In contexts so dependent on mutual expectations, positive theory has great normative influence. If I have strong reasons, based on evolutionary equilibrium selection, to expect you to choose a particular strategy, normative rationality gives me a reason to coordinate with your choice. end p.435 More generally, Skyrms (1996 , 108–9) sees a broadly normative role for evolutionary game theory: What does this all have to do with ethics and political philosophy?I have not said anything about how human beings should live their lives or how society should be organized.There is, nevertheless, a conception of these fields under which this book falls. Ethics is a study of possibilities of how one might live. Political philosophy is the study of how societies might be organized. If possibility is construed generously we have utopian theory. Those who would deal with “men as they are” need to work with a more restrictive sense of possibility. Concern with the interactive dynamics of biological evolution, cultural evolution, and learning provides some interesting constraintsEven those who aim to change the world had better first learn how to describe it.
5.2. What Are the Prospects for Convergence? We have emphasized how recent is our appreciation of the strong connection between rationality and evolution. One difference, of course, is that evolution began as a theory of mindless organisms while game theoretic rationality began as an account of maximally reflective agents. So we close with a pair of contrasting speculations about the effects of our new theoretical insights into the further evolution of human rationality. Both predict that social outcomes will depend on the theory widely held by agents, and this theory will, in part, be selected by social evolution. They differ on the theory recommended. Binmore favors game theory as a normative theory. “If such a theory is ever to do more than grace the dusty shelves of academic libraries, it will be because the combination of education and evolution drives society in the direction of the theory. In this process, the existence of the theory itself will be an important factor. A widely applicable theory of games would, of necessity, involve a strong element of self-prophecy in the sense that the existence of the theory itself would be partly responsible for bringing about stabilizing the event which it `predicts'” (1990, 18–19). McClennen is a critic of the game theoretic account of rationality, which he portrays as an “extremely limiting sort of model” that begs many important questions. So he suggests an evolutionary thought experiment, in which individuals are informed of the possible advantages of more complex forms of decision making, including principle-constrained choice in cooperative dilemmas. (In sec. 3.2 I mentioned how McClennen's proposal might be selected in an evolutionary model using the Ultimatum Game.) “One might then reasonably expect to see this more efficient mode of dynamic decision-making drive out more costly precommitment and enforcement methods, and this through nothing more than what economists like to describe as the ordinary competitive process.” (Danielson 1998b , quoting an earlier version of McClennen 1998 ).
NOTES Sections 2.1 and 2.3 draw on Danielson 1998a . 1. Note that when Sober wrote this in the mid-1980s the tragedy of the commons was still evidently better known than the Prisoners' Dilemma model, which Sober introduces in a footnote. END
