A. Underdetermination

One “pro” consideration invokes concerns about underdetermination. Even if we confine ourselves to normative theories that (claim to) embody liberal democratic values (equal respect for persons and so on), there are a wide variety of different alternatives on offer, including varieties of libertarianism, utilitarianism, and egalitarianism. In my assessment, the largely non-empirical arguments advanced by proponents of each alternative (typically based on supposed a priori insights into what equality, justice, and so on, “really” require) have not been successful in convincing adherents of alternative approaches, suggesting that common commitments shared by the various approaches do not by themselves have sufficiently specific content to single out any particular normative theory as the uniquely correct one. In these circumstances it is worth exploring the possibility that there may be additional constraints that can narrow down the choice among alternative normative theories. Empirical facts about human beings — about human motivation and behavior, about what human beings can know and what they value, about how social and institutional structures influence these interactions — are one possible source of such additional constraints. For example, we may discover that a normative theory that appears attractive in the abstract can only be implemented by means of institutions that have normatively undesirable features or that creates incentives that, given empirical facts of human psychology undermine the values the theory attempts to embody, thereby providing reasons for preferring some alternative theory. From this perspective it is worrisome that current proposals about ideal theory look like attempts to insulate an important part of normative theory from this sort of information. Of course, the persuasiveness of such a project will depend on the details of how it is executed, but to the extent it can be carried out, one would think that normative theorists should welcome it. As discussed below, such a project does not aspire to put normative theorizing out of business — it does not attempt to replace normative theorizing with purely descriptive theorizing about human behavior, or to derive normative conclusions from purely descriptive premises. Instead, in its most plausible version (and as suggested in Section II), it attempts to use descriptive claims, in conjunction with other assumptions, at least some of which will be normative, to constrain the possibilities among (other) normative claims.

One way in which empirical information can be relevant to the assessment of normative theorizing is by using such information to predict likely or possible consequences of implementation. However, another, perhaps more important way in which it can be relevant is the following: often, prior to implementation, we do not know how a normative proposal will work in practice — this will depend on many complex considerations which policy makers will be unable to completely anticipate. (Our theories of human behavior and interaction are too weak for such complete anticipation.) In such cases, there may be no alternative to implementing the proposal in some form and seeing, empirically, what happens — one lets the world provide empirical feedback about the normative proposal. By contrast if we refuse to regard what happens when a normative proposal is implemented as relevant to the assessment of that proposal (on the grounds that the proposal is an exercise in ideal theory and that implementation considerations are irrelevant to ideal theory) we cut ourselves off from this sort of opportunity to learn from experience.

An illustration is provided by Cohen’s response ^{Footnote 22} to Anderson’s charge ^{Footnote 23} that implementation of luck egalitarianism would require “shameful revelation.” Cohen dismisses this as merely an implementation difficulty irrelevant to the correct “conception” of justice — thus ensuring that nothing will be learned about this conception from this and other empirical considerations having to do with implementation. For this to be a reasonable strategy, one would need to be very confident that the correct conception of justice can be identified independently of such considerations, presumably on the basis of non-empirical, a priori considerations of some kind. The track record of a priori theorizing points to the opposite conclusion.

B. The track record of a priori theorizing

Putting aside areas like logic in which precise formal argumentation is possible, philosophical theorizing that aspires to be completely independent of empirical facts —either having to do with human beings or the rest of nature — has, to put it gently, not had a terribly successful track record. (Think of philosophical claims about space and time that ignore physical theorizing about this subject, or philosophical theories of perception that ignore relevant psychology and neurobiology.) Some may respond that while fact-independent strategies may be deficient in the above cases, normative theory (or at least portions of it) is a different matter since it concerns claims about how things ought to be, rather than how they are. However, at least in many areas of inquiry, we don’t think that normative proposals should be pursed completely independently of empirical information. Few nowadays would advocate normative proposals about scientific method that are entirely independent of empirical facts about the behavior of the systems being investigated and the capabilities of those doing the investigating. Theorizing about ideal machines, such as Carnot’s heat engine, which is then used as a normative standard for assessing the performance of actual, physically realizable machines rests in part on empirical assumptions about how nature behaves — Carnot does not, for example, think it worthwhile to consider an ideal machine that violates the conservation of energy. It is not obvious why one should expect matters to be fundamentally different in the moral and political realm. Indeed, as many have observed, a number of the most important contributors to moral and political philosophy (Plato, Aristotle, Hobbes, Hume, Rawls) have made extensive appeals to supposed empirical facts about human beings; so despite current skepticism about (or lack of interest in) such projects, there is plenty of historical precedent for this sort of enterprise.

C. Theoretical versus practical

On one conception of the nature and point of normative thinking — call it theoretical — the goal is to discover what is true or what one ought to believe or judge in the normative realm. In this respect, normative theories are thought to be similar in aspiration to scientific theories — just as different theories of gravitation compete with one another, and individuals try to discover the correct theory, so also for different theories of justice, even if a somewhat different methodology is employed. Assuming this conception, it may be hard to see how, for example, facts about human motivation or what people can know could be relevant to fundamental normative theories. After all, no one thinks that whether people are willing to believe or make use of a theory of gravitation is relevant to whether it is true. This view of normative theory thus leads naturally to the thought that the question of whether a normative theory is “true” should be sharply separated from, for example, whether we can get people to believe it or use it, or what the consequences of their using it might be, and thus to some version of the ideal/nonideal distinction.

I will not try to argue against this “theoretical” conception, but it is important to recognize that there is an alternative according to which the goal or point of normative theorizing is more practical — it involves deciding what to do (where this includes choice of individual actions, but also choices of systems of rules and institutions) and not, completely independently of this, merely deciding what to think or believe. On this conception, normative theorizing in ethics and political philosophy has important features in common with engineering, with choosing a technology or a design, or with solving a prudential problem. Just as it is uncontroversial that in order to design a good automobile one needs to take account of the nature of the materials that go into the car as well as the characteristics of the people who will drive them, so also for the design of moral rules and political institutions — here people and social structures are the relevant material.

One attraction of this more practical conception is that fits well with a naturalistic, non-mysterious understanding of the role that normative thinking about social and political matters plays in the lives of people in general. Why do we have conceptions of justice and related notions at all? What is the point or function of these notions? An extremely natural suggestion is that these (or at least those conceptions that have any claim to normative attractiveness) have developed as a means of structuring and regulating our interactions with one another, so that we can obtain some of the advantages that result from cooperative living while controlling the adverse impact that our behavior can have on others. Sometimes such conceptions and the rules and institutions associated with them reflect explicit and deliberate choices (legislation, constitutions) but more often they are the result of slow trial-and-error processes of social evolution. In either case, they involve patterns of thinking and rules that we human beings construct or create to solve problems in social living. As many have noted, on this constructivist, practical conception, it is completely unsurprising that our thinking about justice and related matters should reflect the influence of “nonideal” practical concerns about implementation, finding schemes that encourage compliance and so on. ^{Footnote 24}

D. Parametric versus strategic thinking

I turn now to another contrast relevant to the ideal/nonideal distinction: the contrast between thinking parametrically and thinking strategically. Any normative theory offering advice about what to do in a situation affecting others will need to take some account of how people behave in that situation. But there are two different ways of attempting to do this. Suppose that one uses some normative theory to attempt to determine (by the lights of that theory) whether to adopt course of action B or B’, which will in turn affect those in group P. A parametric approach recognizes that whether one adopts B or B’ may influence whether those in P adopt action A or A’ in response and includes this fact in the analysis but stops at that point, probing no further. In particular, the decision rule or strategy R1 underlying the choice of A or A’ adopted by those in P is taken to be causally independent of whatever rule, strategy, or principle R2 underlies the choice of B or B’. Similarly the behavior of those outside of P is taken to be independent of R2. In other words, we don’t think about how people (both in A and outside of it) may alter their choices and calculations and the decision rule underlying these (including the range of options considered and how these are weighted) in response to R2, and how this will affect outcomes. Taking such considerations into account involves thinking strategically (and dynamically) — thinking not just about how adoption of some principle may directly affect others, treating their likely behavior as fixed, but how they are likely to modify their behavior in response to the principle adopted, the likely consequences of this, and so on. One thinks in terms of the entire extensive form game, rather than just stopping at the first node. Thinking strategically is an important aspect of nonideal theory and involves considerations that will be missed entirely if one focuses only on ideal theory.

Here are two illustrations. A number of examples in the philosophical literature have the following abstract form: X threatens, credibly, to do something (B1) having very bad consequences unless Y does something (B2) having consequences that are also bad but less so than B1. Focusing only on X’s choice between doing and not doing B1, and the way in which this depends on Y’s choice (and in this respect reasoning parametrically), many writers conclude that “consequentialist” theories recommend Y’s doing B2. One problem with this reasoning is that it is insufficiently strategic and does not take account of various dynamic effects — for example, it does not take account of the likely effects on the behavior of X and others of Y’s adopting this decision rule. Once this adoption becomes known, it follows that X and others can put Y under an obligation to do almost anything, however terrible, by credibly threatening to do something worse. In Singer’s example, the Nazi government could morally obligate people to assist in executions as concentration camp guards by credibly threatening to replace them with more murderous and sadistic guards. ^{Footnote 25} If we take this and other considerations (for example, the desensitizing effect of serving as a guard on Y’s behavior) into account, it is not obvious that (as Singer assumes) consequentialist theories recommend B2 or that they differ from the recommendations of deontological theories in this respect. To the extent that we want to argue for or against consequentialism by considering what it implies about what to do in situations like that described above, excursions into the nonideal (and thinking strategically) will be unavoidable. ^{Footnote 26}

A similar point holds for theories of justice, whether or not they are officially consequentialist — when we attempt to implement these in the nonideal real world, they inevitably create incentives for people to be or to present themselves as being in certain conditions rather than others and to alter their behavior accordingly. For example, attempts to implement a luck egalitarian theory will create incentives for people who want aid to present themselves as (or to be, to the extent they can conceal the role of their choices in this) in conditions over which they have no control — the more they can plausibly deny the role of their own agency, the more aid-worthy they become. Luck egalitarians should consider whether these are incentives they wish to create (particularly given their emphasis on the moral significance of autonomous choice) and this requires thinking strategically. One of the advantages of reciprocity-based considerations is that they are less manipulable in this way. ^{Footnote 27}

A. Definitional moves

It seems uncontroversial that in considering whether to implement a particular set of rules or institutions associated with a normative theory one should take account of how people will behave in the context of those institutions and how this will affect the consequences of implementation. Some theorists respond to this observation by accepting it and (in effect) defining (or redefining) the subject matter of the associated normative theory (or some portion of the normative theory — the “ideal” or “aspirational” part) in such a way as to exclude the relevance of this observation — roughly, by taking the normative theory not to be about implementation at all. For example, Estlund ^{Footnote 28} considers what he calls a Carens market ^{Footnote 29} in which there is a “behavioral requirement” that individuals devote themselves to maximizing their pre-tax income which is then redistributed equally among everyone. Estlund agrees that attempts to actually implement this arrangement would lead to very unsatisfactory outcomes because people will not behave in accord with this requirement. However, Estlund distinguishes the proposal to implement this arrangement, which he calls an “institutional proposal,” and the requirement to implement and comply with it, which he calls an “institutional principle.” He further distinguishes both of these from “fundamental political principles” which he describes as abstract claims without institutional content such as a claim that a certain distributive pattern is just.

Estlund contends that it is no objection to the institutional principle that people should implement and comply with the Carens market that they will not comply with its requirements, since that principle does not say the market should be implemented in the absence of compliance. Estlund associates both sorts of principles — institutional and fundamental political — with the subject matter of ideal theory or “aspirational” theory, to this extent making it true by stipulation or definition (or something close to these) that compliance and implementation issues are irrelevant to such principles and to ideal theory. As noted earlier, a broadly similar strategy is followed by Cohen, who distinguishes between what belongs to the notion of justice, strictly speaking — what he calls the “content of justice” — and to which considerations having to do with implementation and compliance are irrelevant and what he calls “rules of regulation,” which legitimately incorporate such considerations. ^{Footnote 30}

Suppose that we grant, for the sake of argument, that it is possible to introduce the kinds of distinctions described above. As noted in Section II, it is a further and much more controversial claim that these distinctions capture, as an empirical matter, how people think about justice or other moral principles they regard as fundamental — that their thinking is such that they regard all compliance (and other such) issues as irrelevant or “external” to justice. This claim requires empirical evidence, which, as nearly as I can see, Estlund and Cohen do not provide. Moreover, this empirical claim seems dubious, for a number of reasons. ^{Footnote 31} In addition, it would be puzzling (and ironic) indeed if Estlund and Cohen were to be interpreted as arguing for the irrelevance of various empirical considerations about human behavior to normative theorizing in the form of ideal theory on the basis of empirical considerations about humans think.

For this reason, it seems more plausible to interpret Estlund and Cohen as introducing definitional stipulations with no claim that these track actual usage and thinking in any very close way — that is, they should be understood as stipulating that part of what is meant by “institutional principle” or “fundamental principle of justice” or “ideal theory” is that certain empirical facts about human behavior are to be regarded as irrelevant to these. As noted earlier, this raises the question of whether these stipulations are likely to be intellectually fruitful — an issue I explore below.

Let me add that even if the distinctions that Estlund and Cohen introduce are accepted, they have provided no reason to doubt (in fact both seem to agree) that “institutional proposals,” “regulation,” and “nonideal theory” are legitimate parts of normative political theory and, by their own accounts, facts about human motivation and behavior are highly relevant to this enterprise. So one possibility is that the political philosopher interested in bringing empirical facts about human behavior to bear on normative theorizing can meet this portion of Estlund’s and Cohen’s arguments simply by relabeling her recommendations as proposals about institutional principles or regulation. This is one way of understanding my use of empirical evidence about reciprocal behavior in normative theorizing in Section V below, although, as should be obvious from my discussion, I think this concedes far too much to the defenders of ideal theory.

B. The argument from “taint”

There is, however, a more substantive argument that can be extracted from Estlund’s and Cohen’s discussions. Put very roughly, this involves a worry (hereafter, “the argument from taint”) about morally illegitimate compromises intruding into or “tainting” notions of justice or other fundamental moral principles. Imagine a proposed cooperative scheme (for example, the U.S. constitution of 1787) that will succeed in providing crucially important public benefits but only if it incorporates racist elements — the interests and psychology of many of those subject to the scheme is such that if these elements were removed, they would withdraw their support and the scheme would collapse. Others subject to the scheme find these elements deeply morally offensive but acquiesce because they judge correctly that the collapse of cooperation would be worse than its highly flawed continuation. As we ordinarily think about these matters these racist elements should not be regarded as in any way requirements of justice or fundamental moral principles — instead they are compromises with those requirements, acquiesced in because the alternatives are even worse. This shows that not every concession to human motivation that may be required for an institution to deliver some desired result is for that reason alone properly viewed as “required by justice or morality, ” even if on balance the concession is itself justified. Estlund takes considerations of this sort to show that, “it is never a feature’s status as characteristic of humans by nature that constrains the concept of justice, since a judgment is also required in every case about whether the feature’s moral value or significance suits it to have this kind of weight.” ^{Footnote 32}

I accept this claim — indeed, it accords with the schema in Section II for establishing the relevance of factual information to normative conclusions that some connecting normative premise is always required to accomplish this. But what follows from this claim, in my view, is not that empirical facts about “human nature” are always irrelevant to considerations of justice, but rather that if we want to show that such facts are relevant we need to subject them to a certain kind of moral assessment. It is true, of course, that one can always block inferences from factual premises about human motivation and behavior to normative conclusions by assuming the correctness of some favored normative theory N2* and then using this as a basis for excluding as “tainted” all motives and behaviors not conforming to N2* — e.g., we argue that anyone who is not motivated to comply with the requirements of the Carens market is for that reason alone acting from tainted motives which are thus not relevant to what justice requires. I take it, however, that arguments of this sort are plainly question-begging — if one is going to appeal to taint to argue for the irrelevance of various empirical facts about human motivation, some independent argument that such motivations are tainted (or “unreasonable”) needs to be provided. ^{Footnote 33}

C. The positive relationship between empirical information and normative theorizing

How, more positively, might empirical information be relevant to normative theorizing? Let’s distinguish two possibilities. The first focuses on supposed difficulties created by the absence of certain motivational possibilities and behaviors; the second focuses instead on the significance of the presence of certain motivations and behaviors. Arguments of the first sort typically involve claims about either the impossibility of people (or enough people) acting in certain ways or, alternatively, claims about the extreme difficulty or unlikeliness of their doing so. For example, it might be claimed that most people will find it impossible or very difficult to act in accord with the unselfish requirements of the Carens market and that this in itself undercuts such requirements. I agree with critics that such arguments are problematic when taken just by themselves, although, for the reasons described above, I do not find Estlund’s complete dismissal of them persuasive. Even though (let us assume for purposes of discussion) one cannot be morally required to do what is impossible, it is dubious that the behavior and motivation of the sort contemplated in the Carens market (and similar examples) is “impossible” in any relevant sense, at least for many people. Moreover, although the required behavior may be difficult for most people, it can be argued, on the other side, that there are obvious public policy reasons of an incentive-based sort not to allow people to evade otherwise justified requirements by arguing that it is “difficult” or burdensome for them to comply, particularly when the source of the difficulty is “psychological.” ^{Footnote 34}

Consistently with this, it seems to me that, as an empirical matter, in many situations compliance problems arise not so much because those affected are unable to behave in the mandated ways or would find this extremely difficult, but rather because that they think that they are justified in not complying, and regard demands that they do so as illegitimate, particularly in cases in which others are not complying. In other words, people often fail to comply because they think they have morally legitimate reasons for noncompliance. This suggests that rather than resting the case for the relevance of motivational considerations on claims about the difficulty or impossibility of compliance (or of willingness to comply) alone, we should (also) think of empirical considerations as playing a different, more positive role – as providing support for claims about the relevance and importance of certain identifiably moral values and concerns in normative theorizing.

What I have in mind is the following: Suppose one could show that certain motivations and behaviors toward others are very widespread and that these are valued and endorsed by most who exhibit them — that they connect with, embody, of further ends or goals that human beings as an empirical matter care about. Furthermore, suppose one could also show there are no non-question-begging reasons to think that there is a moral requirement not to act on the motives in question. Then one might think that these considerations count as defeasible or prima facie reasons for assigning moral value to the motivations and the action patterns with which the considerations are associated (and the concerns and valuations that underlie these) and to think that it is reasonable and justifiable for them to be incorporated into normative theory in some form. This fits the abstract schema described in Section II, with empirical premises D1 describing human motives and valuations of these, normative premises N1 holding that these are not illegitimate or tainted, and normative conclusions N2 following from these premises. Again, such an argument will not be question-begging or circular as long as the premises N1 are appropriately independent of the conclusions N2. An argumentative strategy of this sort also fits well with the constructivist picture of normative theory described in Section III: at the end of the day, normative requirements can only reflect what we care about or value, suitably filtered by other values we hold — there isn’t some additional, independent source of normativity ^{Footnote 35} .

As an illustration, as a matter of descriptive psychology, most people are strongly motivated to love their children and value their welfare and to act accordingly. Moreover, although there is of course plenty of room for moral argument about just what parents are entitled to do on behalf of their children and how this relates to and trades off against other obligations they have, there are no generally accepted reasons for thinking that some substantial level of concern for one’s children is itself morally illegitimate — on the contrary. The inference pattern I am considering takes these considerations to provide prima facie support for normative theories incorporating these values. The claim that (i) concern for one’s children is not morally illegitimate and has beneficial consequences is not itself tantamount to (and seems acceptable independently of) the claim that (ii) this concern should be reflected or incorporated in any particular way in some normative theory.

The arguments in the following section about the moral significance of reciprocity are intended in roughly the same way. First, empirical evidence shows (unsurprisingly) that people are in fact motivated to reciprocate and value reciprocity in their relationships. Reciprocity sustains patterns of cooperation that provides goods that would not be obtained in the absence of such motives and, in combination with the right rules and institutional structures, protects cooperative schemes against free-riding in a way that other, more unconditional other-regarding motives do not. Reciprocity-based motivations and principles may, as observed above, conflict with other sorts of principles in some situations, but assuming that reciprocity-based considerations should not be dismissed for this reason alone, one might think that there are no obvious independent reasons for treating such considerations as automatically tainted. In particular we need not think of them, for example, as reflecting selfishness or weakness of will.

To enlarge on this last point: the tendency in philosophical discussion of compliance problems, particularly among those who are skeptical about the relevance of compliance issues to normative theorizing, is to focus on cases in which the noncompliance occurs for reasons or motives (selfishness, laziness, stupidity, and so forth) that can be readily portrayed as discreditable. But the existence of reciprocity-based concerns among many people shows that this is a great oversimplification. Suppose our favored theory mandates unconditional aid to others when they meet appropriate conditions, independently of whether one is in relationships involving reciprocity with those others and independently of whether other potential aiders are willing to provide aid. A great deal of empirical evidence, some of it reviewed below, suggests that many people, motivated by reciprocity-based considerations are unwilling to provide aid in such circumstances. But these reciprocators are not selfish in any ordinary sense; they are, after all, willing to contribute or cooperate even when it would be in their self-interest not to do so — they just don’t think that they are required to cooperate if others are not doing so. Someone motivated in this way acts (and thinks of herself as acting in accord with) a distinctive moral set of concerns or values, organized around reciprocity. Note that this is not (or need not be) a matter of the reciprocator finding it psychologically “impossible” or very difficult to comply with requirements of unconditional cooperation. It is rather a matter of the reciprocator having an alternative set of moral beliefs that are at variance with the theory that mandates unconditional cooperation.

As cases of this sort show, compliance issues are not just issues about whether one should comply even if one does not want to or finds it difficult or is too selfish to comply. Compliance also raises assurance issues — normative issues about whether one is morally required to comply when others are not and issues about what to do when the noncompliance of some encourages noncompliance in others. These issues become particularly salient when we recognize that large numbers of people are influenced by reciprocal motives. When a theory of justice produces noncompliance and unwanted results because many people fail to comply for reciprocity-based reasons, we have a problem that is very different from problems that arise when noncompliance results from straightforwardly morally objectionable motives. In cases of the former sort, it is much less easy to dismiss compliance worries as based on considerations that are “external” or “irrelevant “ to core elements of our notion of justice or on “I never said people would be perfect” grounds. ^{Footnote 36}

Once the possible normative significance of the existence of motives to reciprocate is recognized, it is also natural to suppose that more detailed empirical facts about the structure of those motives can also have normative significance. I also develop this idea in the following section.

A. Prisoner’s dilemmas

If an ordinary prisoner’s dilemma (PD) is played only once, and both players are self-interested in the sense of caring only about their own material payoffs, their dominant strategy is of course to defect. If the game is played repeatedly with the same partner and a fixed probability of the game ending in each round, then, if this probability is sufficiently low, cooperative strategies like tit-for-tat (that is, beginning with cooperation and then responding to cooperation with cooperation, and defection with defection) are among the Nash equilibria of the repeated game. This may encourage the thought that self-interest is enough by itself to explain whatever cooperative behavior occurs empirically in repeated games. However, this is dubious. For one thing, there are many other Nash equilibria in the repeated game with self-interested players, including many that are far less cooperative (such as always defect). Appeals to Nash and self-interest do not explain why cooperative equilibria rather than these alternatives so often occur. As an empirical matter, even in one-shot PDs somewhere between 30 and 60 percent of subjects cooperate in typical experiments, with somewhat higher levels of cooperation in repeated PDs. ^{Footnote 37} However, even if a player’s choice to cooperate in a one-shot PD is taken to indicate that she is not purely self-interested, this choice does not distinguish between the hypothesis that the player is a reciprocator and the hypothesis that she is an unconditional altruist of some sort, since versions of both hypotheses predict cooperation. As noted above, a somewhat more sensitive test for the presence of reciprocators is provided by the behavior of second movers in a sequential prisoner’s dilemma. If the second mover (player two) is an unconditional altruist, then, depending on the payoffs and the weight she assigns to player one’s welfare, she may cooperate even if player one has defected. By contrast, if the second mover is a reciprocator, then one would expect whether she cooperates will vary depending on whether player one has cooperated on the first move; player two will reciprocate a cooperative move by player one by cooperating herself, but will be considerably less likely to cooperate if player one has defected.

In a series of experimental tests relevant to the existence of reciprocators, Ahn, Ostrom, and Walker ^{Footnote 38} asked subjects to state their preferences regarding outcomes in one-shot double-blind prisoner’s dilemmas, both involving simultaneous and sequential moves. In both versions, substantial numbers of players ranked (C, C) as more or equally preferred to (D, C) despite the higher material payoff to self in the latter condition, with the preference for (C, C) being stronger in the sequential game. Ahn, Ostram, and Walker also report results from behavioral experiments that are broadly consistent with these expressed preferences: 36 percent of subjects cooperated in the simultaneous move game, and in the sequential game, 56 percent of first movers cooperated and 61percent of second movers cooperated when first movers cooperated. By contrast, virtually no second movers cooperated when first movers defected. The authors also note, however, that these preferences for reciprocation can be strongly influenced by experience: for example, subjects who experience nonreciprocal behavior from others, alter their expressed preferences for reciprocation (even with new partners) and act accordingly.

An important feature of the experimental results described by Ahn, Ostram, and Walker. (which also appear in virtually all other studies of reciprocity and trust) and to which I will return below is the apparent heterogeneity of motivation across different subjects. In Ahn, Ostram, and Walker’s study, although roughly 60 percent of second movers cooperate when first movers cooperated in the experiment, this leaves approximately 40 percent who behaved in a self-interested way in this condition. Other experimental results support a similar picture, according to which most populations consist of a mix of different types or varieties of subjects, some of whom seem to be strongly motivated by reciprocal considerations (when these are relevant), and others of whom seem to be motivated mainly by self-interest. A third, smaller group seems more strongly motivated by nonreciprocal, nonconditional other-regarding motives even in circumstances in which considerations bearing on reciprocation seem relevant — for example, they will continue to cooperate even in the face of defection by their partners. As we shall see, this heterogeneity has important implications for issues having to do with the circumstances under which cooperative schemes will be undermined by noncompliance. Reciprocal strategies are particularly attractive under such circumstances, but again to recognize this we need to think nonideally.

B. Public goods games

In a linear public goods game, each player i from a group of N players chooses an amount c _i (from an initial endowment w _i which is the same for each player) to contribute to a public good. The sum ∑c _i of these individual contributions is multiplied by some factor m (1/N <m <1) and this total is then divided equally among all of the players, regardless of how much they contributed. This game may be either played once or repeated, with total contributions announced after each round. In both cases, the dominant strategy for a purely self-interested player is to “free-ride,” receiving his share of the contributions of others, but contributing nothing. Thus, in this case whatever cooperative behavior occurs cannot be understood as a repeated game effect involving selfish players.

Experimental studies of behavior in repeated public goods show that average contributions across all subjects begin at a relatively high level (around 50 percent of total endowment), but that there is considerable variation in individual contributions, with some subjects contributing nothing from the first round onward. When the game is repeated, average contributions typically decline sharply over time.

The most plausible explanation for this decline is that the subjects are a mix of self-interested players and reciprocators. Reciprocators begin by cooperating in the early rounds of the game, expecting or hoping that others will as well. By contrast, self-interested types contribute nothing from the start. The decline in contributions occurs because reciprocators gradually stop contributing in response to the noncontributing behavior of self-interested players — reciprocators are not willing to continue to contribute while others free-ride. This interpretation is supported by subject’s verbal explanations of their own behavior as well as a number of other pieces of evidence including “restart” and “grouping” effects.

In the restart version, a public goods game is repeated for, say, ten rounds, stopped, and then “restarted” in the sense that subjects are told that a new ten-round repeated game will begin with the same players. ^{Footnote 39} Contributions exhibit the standard pattern of decline over the first ten rounds, but the restart temporarily boosts contributions. This behavior makes sense if many of the subjects are reciprocators who are willing to contribute when the restart begins as part of an attempt to signal anew their willingness to cooperate if others are willing to do so as well. By contrast, this behavior makes little sense if the subjects are purely self-interested and just contribute in early rounds out of confusion, as some have claimed. We would also not expect to see this pattern of decline in contributions over time if the subjects who are not self-interested are unconditional altruists or if the motivations are some weighted average of self-interest and unconditional altruism — the contributions of such subjects would remain stable over time or even increase, as contributions from others decline.

Experiments involving grouping effects also support this interpretation of public goods games as involving interactions between reciprocators and more self–interested players with declines in contribution, when they occur, deriving from the adverse impact of the latter on the former. A representative example is an experiment conducted by Page, Putterman, and Unel ^{Footnote 40} in which subjects were given information every three rounds about the contributions of other subjects in previous rounds and were also given an opportunity to express preferences about future partners. Subjects whose mutual rankings were the lowest (with a low ranking meaning that a partner is most preferred for cooperation) were then organized into a first group; those whose rankings were intermediate were organized into a second group, and so on, for a total of four groups. The result of this “endogenous grouping treatment” was that the average level of contribution across all groups (70 percent) was much higher than the average level of contribution in a baseline treatment in which there is no regrouping (38 percent). Moreover, average levels of contribution varied greatly across the groups, with this following the order of group formation — that is, contributions were highest and continued to be so in the first group, formed of the highest cooperators, and were lowest in the fourth group: in the first group, 50 percent contributed their entire endowment in the final period (with the corresponding numbers from the second, third, and fourth groups being 43, 18, and 0).

Another relevant piece of evidence is this: if a significant portion of subjects are reciprocators, one would expect their contributions to be positively correlated with their beliefs about the extent to which others are contributing, since reciprocators contribute more to the extent that they believe others are contributing more. This prediction contrasts with the prediction that follows from other hypotheses about the motivations of subjects who contribute. For example, a simple “utilitarian” model according to which subjects are unconditional altruists who care, in addition to their own self-interest, only about whether some optimal aggregate amount is contributed in toto, predicts that individual subjects will contribute less if they believe that others will contribute more, and will contribute more if they believe others will contribute less. In an experimental study employing a repeated public goods game, Croson found, in accordance with the hypothesis that a large portion of subjects are reciprocators, a strong positive correlation between a subject’s beliefs about how much others would contribute and that subject’s actual contributions, as well as a correlation between individual contributions and the actual level of other’s contribution across different periods of the game. ^{Footnote 41} Many field studies show a similar pattern—for example, people’s willingness to fully pay taxes is highly correlated with their beliefs about whether others are similarly complying. ^{Footnote 42}

C. Punishment

A number of recent experiments have explored the effects of introducing a costly punishment option into a repeated public goods game. ^{Footnote 43} This allows subjects, at cost c to themselves, to punish noncontributors by imposing a fine that reduces their earnings by amount r where r>c. Under this condition, a number of subjects punish noncontributors, even in the final round of the game, in which punishment cannot influence future behavior, so that punishment is clearly not a strategy that increases the punisher’s expected payoff. Introduction of this option prevents the decline in contributions with repeated play that is otherwise observed. Again, an obvious interpretation is that punishers are reciprocators who are willing to engage in negative reciprocation, even when it is not in their self-interest to do so. In the absence of an explicit punishment option, reciprocators have no way of responding negatively to free-riding except by withdrawing their contributions; when a punishment option is provided, they make use of this instead and do not reduce their contributions.

D. Trust games

McCabe, Rigdon, and Smith ^{Footnote 44} compared behavior in following two extensive form games (both variants on “trust games”). ^{Footnote 45}

In Figure 1, player 1 moves first, at choice node 1. Player 1 can choose to move across, ending the game and ensuring a payoff of (20, 20) (first number is player one’s payoff) for herself and player two. Alternatively, player one can move down, in which case player two has a choice of moving across, with payoffs (25, 25) or moving down with payoffs (15, 30). Player 1’s choice to move down is thus a “trusting” move in the sense that she foregoes a certain payoff of 20 in order to give player two an opportunity to choose an alternative that is better (25, 25) for both, but also trusting that player two will not choose down, producing an outcome that is still better for player 2 (30 versus 25), but worse for player 1. Put slightly differently, in choosing down, player 1 confers a benefit on player 2 which she expects or hopes will be reciprocated by player 2’s choosing in a way that benefits player 1. To the extent that player 1 chooses in this way, this is an example of the use of the conferral of benefit to induce reciprocation. A natural hypothesis about interactions of this type, described by Philip Pettit ^{Footnote 46} is to suppose that under the right conditions the trust communicated by Player 1’s choice of down can help to induce player 2 to behave reciprocally and that when Player 1 moves down, she acts with this intention or expectation.

Figure 1:

Trust Game With Outside Option

Now compare this with the game in Figure 2, which is just like the game in Figure 1, except that the outside option has been removed — now player 1’s only choice is to move down. Because of this, player 1’s move down cannot communicate to player 2 an intention to trust or help player 2. Supposing that something like Pettit’s analysis is correct, this removes one of player 2’s motives to move across: the motive of reciprocating the trust shown by player 1. As a result, one would expect that to the extent that player 2 is the sort of reciprocator who responds to the intentions with which player 1 acts, player 2 would be more likely to choose down in the second version of the game than in the first version. As the numbers attached to the game trees (which give the proportions of players choosing each alternative at each node) this is indeed what is found — the difference in choices of down versus across by player 2 between the two versions of the game is significant at the 0.01 level. Although, as we see from Figure 2, a significant number (33 percent) of second movers are influenced by a nonreciprocity-based or unconditionally altruistic desire to benefit player one (at a cost to themselves), even when there is no prior provision of benefit by player one, the increased proportion of second movers who move across in the first version of the game (65 percent) strongly suggests that a powerful additional, independent motive to reciprocate is operative when a prior benefit has been provided. Moreover, first movers in the first version of the game (accurately) anticipate this response, and this expectation plays a role in motivating their choices.

Figure 2:

Trust Game Without Outside Option

E. Normative significance

If we were to incorporate these observations into a normative account of the role of reciprocity (justice as reciprocity) proceeding along the argumentative lines sketched in Section V, we are led to the following picture. First, there are a distinctive set of motives and associated normative concerns that are distinct both from self-interest and from unconditionally altruistic motives (whether these are utilitarian or linked to the achievement of some favored distribution pattern that is not conditional on people’s willingness to reciprocate), just as Gibbard claimed. Although this is not the place for extended discussion, this suggests a number of points concerning the design of distributive schemes. For those of us favoring generous social welfare programs, it is important to design such programs so that they are supported by reciprocal motives rather than conflicting with them: Social Security in the United States is an example of the successful recruitment of such motives, while, as argued by Fong, Bowles, and Gintis, many programs that provide aid to the disadvantaged would probably attract stronger support if they more clearly conformed to moral values built around reciprocity, and avoided, when doing so is possible and consistent with other values, forms of aid that are unconditional in the sense of requiring no form of reciprocity. ^{Footnote 47} On the other (“advantaged”) side, a concern with reciprocity furnishes a powerful argument — one, unlike most philosophical theories of justice, that appeals to considerations actually invoked in political discourse ^{Footnote 48} — that those whose economic success depends upon infrastructure and cooperation provided by others have obligations grounded in reciprocity to “give something back” to others.

Second, results described above (for example, the extensive form version of the trust game) suggest that people treat situations in which others are worse off and can be aided but in which no history of reciprocation exists very differently from otherwise similar situations in which reciprocal relationships exist. They are more willing to provide — and more likely to think that they ought to provide — aid in the latter case. At the same time, however, motives to reciprocate are not the only other-regarding motives people possess — motives of a more unconditional sort are also present. This suggests a normative theory in which the requirements of distributive justice and the provision of aid are at least somewhat different for those with whom we are involved in ongoing relations of reciprocity (for instance, inhabitants of our neighborhood, city, country) than those (if any) with whom we are not so involved — a feature with obvious implications for theorizing about justice in international contexts. Third, as the trust game example and numerous others illustrate, reciprocators care about the intentions with which others act when reciprocation is possible and not just the magnitude of the material benefits involved. This suggests that an adequate normative theory that incorporates notions of reciprocity should not just focus on the material benefits obtained through the reciprocal relationships but also on what reciprocators express about their attitudes toward each other through their interactions, since this is reflected in the intentions with which they act. To the extent that reciprocity matters in justice and elsewhere in normative theorizing, this points us toward a relational conception in the sense of Anderson ^{Footnote 49} according to which these have to do with the establishment and regulation of relationships among people and not just the realization of particular patterns of distribution of material goods. In particular, reciprocity is characterized by relationships among people that involve the absence of exploitation of other’s contributive behavior, free-riding, “taking without giving back,” and so on. Although I will not argue for this claim here, I see this is a potent source of criticism of many of our present social and political arrangements.