Evolutionary game theory and the normative theory of institutional design: Binmore and behavioral economics

Don Ross[1]

University of Alabama at Birmingham

University of Cape Town

dross@commerce.uct.ac.za

Abstract

In this paper, I critically respond to Herbert Gintis’s criticisms of the behavioral-economic foundations of Ken Binmore’s game-theoretic theory of justice. Gintis, I argue, fails to take full account of the normative requirements Binmore sets for his account, and also ignores what I call the `scale-relativity’ considerations built into Binmore’s approach to modeling human evolution. Paul Seabright’s criticism of Binmore, I note, repeats these oversights. In the course of answering Gintis’s and Seabright’s objections, I clarify and extend Binmore’s theory in a number of respects, integrating it with Kim Sterelny’s and Don Ross’s recent (respective) work on the evolution of people as cultural entities. The account also yields a novel basis for choosing between socialism (broadly conceived) and what Binmore calls `whiggery’ as normative political programs.

Biographical sketch

Don Ross: Ph.D., University of Western Ontario 1990. Currently Professor in the Center for Ethics and Values in the Sciences, University of Alabama at Birmingham, and Professor of Economics, University of Cape Town. Author or editor of 8 books and about 40 articles on the foundations of the behavioral sciences, game theory, economics and general philosophy of science. Chief economic consultant to the South African road pavements industry. The first volume of his two-volume Economic Theory and Cognitive Science will appear in June 2005 from MIT Press.

Keywords

Theory of justice; bargaining theory; evolutionary game theory; human evolution; Ken Binmore; Herbert Gintis; Kim Sterelny

Are there objectively true answers, in general, to the following two questions: (1) What utility function is (particular) person P maximizing in strategic situation S?; and (2) What game are persons P, Q, …, R playing in S? Ken Binmore, in Natural Justice,[2] supposes that these questions indeed have factual answers, even if (of course) an analyst’s evidence might not always permit her to find them. Herbert Gintis, in his discussion of Binmore’s book in this journal,[3] also supposes that the questions have factual answers, but he thinks that people typically play different games, and typically have different utility functions, than Binmore does. As a result, Binmore and Gintis disagree about the correct explanation of stable social conventions and the basis of people’s normative social judgments, and about the kinds of social policies that will prove sustainable and produce (desirable) intended consequences. They don’t disagree over what is desirable: efficient dynamic equilibria that tend to take us toward progressively more egalitarian distributions. Both deny that we face any systematic trade-off between efficiency and equality.

I will argue here that both Binmore and Gintis regard the basis for objectivity of answers to questions (1) and (2) a bit simplistically. This will certainly not be an argument for the conclusion that nothing we can factually know is relevant to answering the questions; like Binmore and Gintis, I believe in naturalism, that is, in the thesis that science should guide our morals and our public policy. And there is no sense to be had in talking about science without objectivity. However, I will argue that if all the facts were in about any particular run of human behavior, for any particular set of interacting people, we would still face some decisions in answering questions (1) and (2) that would have to be based on pragmatic considerations. I will then argue that these pragmatic considerations, where normative inquiry is concerned, favor Binmore’s general way of answering the questions. There will be no discussion of first-order moral theory, because I agree with Binmore and Gintis about what is socially good. There will be some brief remarks on meta-ethical issues at the end.

I take the significance of the discussion to reside partly in the fact that Gintis, in light of the network of collaborators with whom he works, can be read as representative of the behavioral economists who are criticized at several points in Binmore’s book. Binmore, for his part, has for over a decade been applying evolutionary game theory to normative issues that have been central in political philosophy for the past half-century and more. He has repeatedly argued that much of this philosophy has been inadequate because it has incorporated an empirically implausible model of human motivation. Gintis has been similarly critical of recent thinkers, on both descriptive and normative topics in human behavior, who have not been enlightened by evolutionary game theory. To the extent that we find Binmore and Gintis in disagreement on some issues, then, we are motivated to investigate their disputes out of concern for the unification of descriptive behavioral science and normative political theory.

The discussion will be organized as follows. First I will describe Binmore’s view of the answers to my opening questions. Then I will describe Gintis’s answers as these contrast with Binmore’s. Then I will say what I think is the correct view, and why. Finally, I will give the reasons why the correct view produces a policy theory closer to Binmore’s than to Gintis’s.

1 Binmore on utility and game individuation

Like all serious economists since the 1930s, Binmore repudiates sensationalistic interpretations of utility. A utility function for an agent is simply a description of that agent’s behavior in a form that makes the behavior amenable to treatment by means of the generalizations of economic theory. `Utility’ denotes properties of schemes of representation, not properties of any putative psychological states. Why is description in terms of utility functions necessary for economic generalization? The answer is that economic reasoning relies on solving maximization problems. We do not need to suppose, in applying economics to agents, that they always succeed in behaviorally solving their maximization problems as economists identify them. People, Binmore argues[4] often fail to maximize because they need to learn the structures of their environments, including their social environments. Thus they sometimes behave in games in ways that economists should regard as errors relative to the utility functions we have assigned. For example, Binmore believes that people will tend to cooperate in one-shot Prisoner’s Dilemmas when first confronted with them because such games are unusual in light of their experience[5], so they will initially play them as if they were rounds in repeated games. However, empirical evidence (as surveyed by Ledyard[6]) indicates that as people gain experience in one-shot PDs, they learn to play Nash equilibrium (NE) strategies. That is, they learn to choose strategies that maximize their utility within the set of available NE – which in the case of the one-shot PD of course means the unique, Pareto-inefficient NE of mutual defection.

This implies that the behavior described by a utility function can’t just be the run of actually observed behavior. Otherwise, as people learn to modify their behavior the analyst would have to constantly adjust their utility functions. Naïve players of one-shot PDs would have to be assigned preferences for cooperative outcomes, while sophisticated players would be assigned preferences for outcomes that are `selfish’ relative to the first sort. In fact, as long as we stuck carefully to the revealed-preference interpretation of utility, if utility functions simply summarized observed behavior then we wouldn’t be able to represent Binmore’s naïve players as involved in one-shot PDs at all, since we’d have to assign them utility functions such that their payoff matrices would yield different games.

This is not how Binmore proposes that we build our models. Utility functions for him must describe behavioral dispositions. The dispositions in question must be identified by reference to equilibria on which behavior converges within the time scale (short, medium or long run) of the class of games under analysis. But how could we know, in the case of naïve players, where their pattern of play will eventually converge? The answer can only be that we know, from observing other, sophisticated, players who serve as models of the naïve players, that the naïve are presently playing out of equilibrium.

This in turn implies a certain evolutionary-psychological theory of people. Binmore is explicit that he does not think evolution has built them to be myopic maximizers of wealth[7]. Thus we cannot necessarily suppose that two people have jointly achieved Pareto optimality if they achieve the cooperative outcome (say, through external enforcement) in what would be a PD if only their potential monetary rewards were considered; the players may have preferences at equilibrium over goods besides money (especially, Binmore emphasizes,[8] for social status). Nor should we assume that people are always or usually myopic maximizers of their biological fitness. Cultural selection, Binmore argues, may have conditioned them to have `expanded their moral circle,’[9] so that they are disposed in equilibrium to treat unrelated people as though they were kin. Such people’s behavioral dispositions at equilibrium will be described by utility functions that favor the fitness of genes that are not parts of their own genotypes.

Thus it is not analytic for Binmore that people are, in Gintis’s terms, `self-regarding’. One could coherently model processes of cultural evolution as sequences of evolutionary games that condition people to behave like ants, in contexts of very wide moral circles. Binmore thinks it is just empirically false that this describes human history. Borrowing a style of reasoning from evolutionary psychologists[10], Binmore argues that cultural evolution has had time to produce norms that help people coordinate on equilibria in repeated games, but hasn’t had time to re-wire biological preference structures that are described by Hamilton’s equation for kin selection. This is of course speculation: we have no consensus quantitative model that tells us how fast cultural evolution can work. However, it is based on empirical evidence, both systematic, as in the case of the experimental data on subjects playing PDs, and homely, as in the following observations that are surely hard to argue with: “Apart from a few saints, who gives so much of their income to charity that it really hurts? Who would keep on plunging into the river if there were always a stranger struggling for his life when you took the dog out for a walk?”[11].

Thus Binmore’s economic model of people can be summarized as follows. A typical individual person biologically inherits some dispositions to act in accordance with the standard utility function of a sexually reproducing, diploid organism. But we can then distinguish at least four importantly different types of scenarios in which these dispositions are manifest:

(1) In a culturally familiar repeated game, people who share a culture will tend to coordinate efficiently (in accordance with Nash bargaining) on one of the equilibria. Their beliefs about what is fair will simply express this coordination history; moral norms just are institutional devices for stabilizing coordination.

In novel games, people will mistakenly apply these norms and thus tend not to choose equilibrium strategies at first. So,

(2) Where a novel game is relatively simple and doesn’t raise equilibrium selection problems, as in a one-shot PD, people will quickly learn that the moral norm serves them poorly and will drop it. They’ll converge on NE strategies. (Thus, a researcher won’t be able to get someone into a one-shot PD over money with their own mother, because Hamilton’s equation will rule out the relevant payoff matrix; but two strangers will be induced to defect following a period of learning.)

(3) Where people have mistakenly modeled one another as relatives – say, in an army platoon or a small religious cult – they’ll also play a different game from the one they’d play with strangers in the same situation. Foxhole buddies won’t readily get into PDs, just as people and their mothers won’t.

(4) In novel repeated games with multiple equilibria, or familiar such games with representatives of other cultures, people will also apply their familiar norms and get non-equilibrium outcomes at first, and typically for some time. A great deal of instability and inefficiency, perhaps including much violence, may go on. But eventually, if genocide or enforced segregation doesn’t eliminate the novel games, then moral norms will themselves adjust by cultural evolution until coordination on new equilibria is achieved. The new norms will be ready for misapplication to the next round of novel games that are generated by technological change, migration, or drift represented by subcultural speciation.

I want to later be able to compare Binmore’s and Gintis’s models of people as economic agents. To facilitate this, I’ll need to describe their respective interpretations in terms of a distinction neither of them draws explicitly, but that can be found in Clark[12] and Ross[13]. This is the distinction between a biological individual (BI) and a person. It is drawn by reference to, and for the purpose of, economic analysis in which the basic units, whatever they are, must be identified with utility functions. In what Binmore calls the `long-run’ dynamic, the players of evolutionary games must be strategies that compete to maximize fitness. They will be represented in each particular round of the long-run game by transient entities whose utility functions must be modeled according to Hamilton’s rule. These entities are BI’s. From the long-run perspective, their cultural and personal idiosyncrasies – do some but not others approve of polygamy? engage in suicide bombing? go down with the ship? – must be regarded as noise. Now, what is noise at one level of abstraction is often data at another. In the medium run, cultural differences are relevant to equilibrium properties, and in the short run personal differences are relevant. People, entities that stabilize themselves into consistent and distinctive characters under pressure from others with whom they have to try to coordinate, are analytically separate entities from the BI’s with which they are materially coextensive because they have different utility functions. In assignment of utility functions, we confront a relativity of scale that is familiar in most or all sciences[14].

Accepting scale relativity to facilitate analysis – in this case, economic analysis – of course doesn’t imply accepting it as a brute limitation on explanatory unification (as fallaciously maintained by Dupré[15]). Though a BI can be distinguished for modeling purposes from the person with whom she is coextensive, these entities are obviously related in an intimate way, and unification of behavioral science partly consists in stating the relevant relation. Much of the disagreement between Binmore and Gintis, we shall see, turns on different interpretations of this relation. Furthermore, I will argue, such interpretation is empirically underdetermined. This does not imply that we must regard it as free or arbitrary.

Partly because Binmore is not explicit in distinguishing between BI’s and people, he doesn’t provide a game-theoretic account of the evolution of people in a world that was once inhabited only by BI’s. Ross[16] attempts this, in a way that will be summarized later. However, Binmore does offer a game-theoretic story about the evolution of the moral conventions that partly distinguish people from BI’s. Indeed, telling and justifying this story is the principal activity of his book. (This should be distinguished from the book’s main purpose, which is to advise us on how to avoid political-institutional disasters given evolved norms characteristic of modern secular societies – so, Britain, Boston or South Africa, but perhaps not Afghanistan or Alabama.)

Here, then, is a summary of the life cycle of a moral norm according to Binmore. Thanks to kin selection, evolution can build social BI’s who are disposed to cooperate with relatives. In species confronting unstable environments,[17] achievement of cooperation will require BI’s to cognitively model one another’s dispositions with reasonable accuracy. They thus evolve empathetic preferences, that is, models of one another’s relative well being that are indexed to each individual’s conception of their own good. Thus equipped, they can bargain to equilibria in which mutual gains are realized. The bargaining equilibria in question will reflect differences in power. However, their most important games are all repeated ones, and so the folk theorem of repeated game theory tells us that they face (in the abstract) equilibrium selection problems. Accidents of history will generate focal points that, over time, solve these problems by serving as anchors for the stabilization of conventions. The conventions in question tell each individual what constitutes equilibrium behavior in her society. In particular, they tell her when she can expect favors to others to be reciprocated, and to what degree, and when, if a favor is not reciprocated, she can call on others to punish the free rider. Since conventions must tell everyone what to do in every social role they’ll typically occupy at some time, they thus also tell the individual when she is obligated to reward and punish others herself, and when she should expect punishment if she yields to what the community defines as selfish temptation.

None of this yet requires invocation of the kinds of distinctively moralized appeals we find made by people, but not by other animals (as far as we can tell[18]); the conventions could all just be implicit in behavior, and transmitted by copying, as in the models of Young[19]. However, two considerations – only one of them explicitly noticed by Binmore – imply something more. First, Sterelny[20] draws attention to the fact that human conventions apply to classes of logically similar situations that are not reliably marked by stereotypical perceptual salience cues. Thus people’s ability to swiftly and confidently apply conventions to novel situations shows that a natural cognitive endowment of some sort allows them to `decouple’ their cognitive models of the conventions from specific, stereotyped behavioral responses. Institutional moral conventions can evolve to encode these more abstract relationships. Second, as Binmore but not Sterelny appreciates in so many words, explicit fairness norms can enable people to coordinate on equilibria in novel situations. These will be the kinds of situations in which people notice and explicitly invoke their own norms. When they deploy them in familiar scenarios, they’ll tend to be unaware of doing so. Thus, as Binmore puts it, medium-run use of a body of fairness norms will tend to leach out the norms’ moral content and reduce them to purely habitual judgments about the relative status of different sorts of people. Here lie the roots of the sort of corruption that Binmore’s work aims, as a normative project, to upset: people can fail to notice when changing circumstances make better equilibria available to them, because they go on playing habitually rather than re-packaging their cultural salience landscapes in light of new feasibility sets. Indeed, some people who enjoy privileges that others could strip away while still staying on equilibrium paths will be incentivized to actively suppress general recognition that new fairness norms are feasible; these people represent each generation’s version of conservatism. Binmore’s work is intended as a weapon against conservatism that at the same time avoids utopian disdain for the importance of stability (that is, for the importance of keeping institutional conventions on equilibrium paths).

Thus, for Binmore, we answer our two opening questions, in any given instance, by representing the recent cultural-anthropological history of the people whose behavior we want to explain – and reform – in game-theoretic formalism. The utility functions of these people – these cultural artifacts – will not be the utility functions of their ancestral BI’s. However much his rhetoric might sometimes give the opposite impression, Binmore does not follow Tooby and Cosmides[21] in seeing people as instantiations of Pleistocene hunter-gatherer utility functions adrift in and maladapted to modern industrial societies. Medium-run dynamics are given more transformative weight by Binmore than they are by the kind of evolutionary psychology that left-wing humanists (e.g., Dupré[22]) often deride. The games people will play will be determined by reference to the preference structures their behavior would reveal in medium-run equilibrium given learning over the short run. Thus, for Binmore, no round of a game in which people cooperate at medium-run equilibrium can, as a matter of logic, be modeled as a one-shot PD. This is a function of how we should use our mathematics to describe reality; it is not an empirical claim about the extent to which modern people do or don’t resemble their ancestral BI’s or their Pleistocene forebears. Binmore has his own opinions about these empirical matters, but they are logically separable from his views about how we ought to build game-theoretic models.

2 Gintis on human cooperativeness

Gintis introduces his core disagreement with Binmore as follows: “by contrast with Binmore, I do not believe … moral rules involve applying local cultural social indices to the Nash bargaining problem … Rather, I believe that human beings are emotionally constituted, by virtue of their evolutionary history, to embrace prosocial and altruistic notions of ingroup / outgroup identification and reciprocity. These aspects of human nature are incompatible with Binmore’s notion that humans are self-regarding creatures.”[23] Gintis says that this disagreement “impact[s] upon the social policy implications of ethical theory.” Unfortunately, he does not clearly identify these potential impacts. I will therefore offer some speculative arguments about them later.

The heart of Gintis’s objection to Binmore lies in the former’s claim that people’s preferences are not self-regarding. By this he means that “people care about payoffs to others, not just themselves, and individuals care about how outcomes are generated, not just the outcomes themselves.”[24] As Gintis notes, this raises no problems for mainstream utility theory, which requires only that agents’ preferences be transitive and complete (at least, across any well justified domains of application of the model). “The notion that agents maximize utility does not require that agents be self-regarding,” Gintis says, “since there is no connection between the transitivity of preferences and the content of preferences. Indeed, one can apply standard choice theory, including the derivation of demand curves, plotting concave indifference curves, and finding price elasticities, for such preferences as charitable giving and punitive retribution.”[25] Now, as noted above, Binmore does not believe, as an empirical claim, that people are narrowly selfish in the `traditional’ economist’s sense[26]; he thinks that most modern people operate within `expanded moral circles,’ which means that they behave according to norms that lead them to treat many non-relatives as though they were relatives. As we will see later, this means that we can treat people as self-regarding insofar as their agency is dynamically consistent with that of their ancestral BI’s.[27] But since Binmore derives people’s utility functions by reference to their behavior in medium-run equilibria following evolution of fairness norms, many of the preferences Gintis calls `prosocial’ will appear as arguments in most people’s utility functions according to Binmore too. Thus it is relevant to both of them, and not just to Gintis, that the formal framework of microeconomic theory is sufficiently flexible in its representational power to express their views of people without requiring foundational revision.[28] It thus seems at first glance that their dispute must be entirely and directly empirical.

I aim to argue that, this appearance notwithstanding, the argument has empirical aspects, but is mainly methodological. The issue is confused, I will maintain, by the fact that two orthogonal methodological differences are implicated, which neither Binmore nor Gintis are consistently clear about. Sometimes philosophy really can help empirical inquiry, even if Kantian philosophy, either moral or epistemological, never does, as both Binmore and Gintis rightly insist. But before I can cleanly pursue this argument, we need to get all the points of overt disagreement out on the table. Let us construct a list, ordering it by reference to the sequence of objections in Gintis’s paper.

Gintis argues that Nash bargaining over equilibria within feasible sets cannot reliably yield stable coordination around particular fairness norms in large groups (such as modern political communities like states). Binmore’s whole approach is built around the assumption that such coordination can be stable.[29]

Binmore predicts that learning will cause people’s behavior in one-shot games to converge toward narrow selfishness, at least where they are not close relatives or foxhole buddies. Gintis reviews experimental evidence that such convergence does not happen, at least in ultimatum and dictator games among small groups.[30] (See Henrich et al[31] for the latest and most carefully accumulated evidence of this sort.)

Gintis maintains that in one-shot games with strangers, people should be modeled as successfully maximizing utility functions. Binmore resists this modeling approach, on grounds that it leads to vicious underdetermination of theory by evidence. Says he on this point: “I would be discouraged to find that each new experiment seems to require subjects to have a brand new utility function, but behavioralists are made of sterner stuff than me!”[32]. For Binmore, utility functions should be constructed on the basis of preferences revealed in repeated games, or one-shot games with which subjects have been made thoroughly familiar using valuable incentives. Thus Binmore doubts that naïve subjects really play the one-shot PDs, ultimatum games and dictator games that Gintis takes them to be playing when he interprets experiments. In these cases, Binmore and Gintis give divergent answers to the two questions that opened this paper.

Binmore predicts that people will regularly engage in second-order punishment of people who fail to engage in first-order punishment of norm-violating free riders. (That is, a first party observing a second party failing to punish a third party who deserves punishment according to prevailing norms, and where the norms assign punishment responsibility in the circumstance type to the second party, will punish the second party.) Gintis claims that this prediction is confuted by evidence.

Gintis thinks that his framework directly explains why people support policies generous to the poor when they deem the poor to be innocent victims, but not when they deem the poor to be substantially responsible for their own conditions. He does not directly say that Binmore’s framework cannot explain this, but he says that it “cannot be explained away by a fuller and more rigorous account of self-interest,”[33] by which I infer allusion to Binmore’s account from the fact that he includes the issue in his criticism of Binmore’s work.

Gintis claims that people do not generally “transform their emotional responses when they are in conflict with their material welfare,”[34] whereas Binmore predicts that they do. (This is the process by which, for Binmore, moral content is leached out of fairness norms by cultural habituation.)

Of these disagreements, (3) is the one that is obviously methodological, even if it is sensitive to empirical issues (as every non-empty methodological disagreement had better be). This is why I opened this paper with the questions I did: I aim, as far as possible, to reduce the other disagreements to this one, and then offer a pragmatic defense of Binmore’s position on it. (This will require some minor amendments to Binmore’s view, however; Gintis’s arguments do require some concessions.)

3 Scale relativity in modeling people and games

Among the disputes enumerated above, (1) is the most directly theoretical. Gintis defends his side of it mainly (though not exclusively) by considering properties of abstract models. A review of models shows that in large n populations where some information is private – thus introducing noise about who deserves reciprocation and punishment and when – there is path-dependency: special restrictions are required to yield dynamics in which equilibria, once lost, will be recoverable[35]. Folk theorem models, on which Binmore relies, typically generate stability by invoking use of trigger strategies[36] by all agents; however, trigger strategies are insufficiently discriminating to preserve coalitions given non-tiny n and imperfect information. They can yield stability, but only at low levels of cooperation. This problem can be avoided if agents resort to more discriminating punishment strategies. However, for this to work either punishment costs must be low, or agents must deploy second-order punishment strategies. This leads to issue (4) above, where Gintis claims that Binmore is empirically confuted.

I find nothing with which to quibble in Gintis’s interpretation of the models he discusses. We need not, however, accept without demur his assumptions about how the predictions of the model correspond to empirical phenomena. I will concentrate on two problems (though I do not think these are the only ones).

The most interesting problem, in my view, has to do with punishment costs. There are good reasons why most of the models Gintis reviews assume non-trivial costs to use of discriminating (as opposed to generalized trigger) punishment strategies. One reason is that many of the models have their historical roots in the study of industrial cartels. Here, because markets can very efficiently select strategies for tiny relative advantages in profitability, Cournot and Bertrand equilibria (depending on which game parameters are free) are strong attractors. This is equivalent to saying that agents face non-trivial opportunity costs from using non-discriminating punishment: a firm that altruistically defends the cartel against free riders automatically cedes some market share to its other rivals[37]. A second (and deeper) reason, which will be pursued for the rest of this section and into the next one, is that models in evolutionary game theory have typically taken the basic empirical phenomenon under investigation to be the initial evolution of sociality in populations of asocial organisms. In this context, it is indeed hard to imagine circumstances in which one organism could significantly harm the welfare of another without using non-trivial energy resources or running non-trivial risks.

Notice, however, that there is no disagreement between Binmore and Gintis with respect to that problem. They agree that sociality must have arisen through the co-opting of mechanisms that support kin selection. Now, evolution has, roughly speaking, stumbled across two paths for ratcheting relative-discrimination into sociality. One path has been followed by the eusocial animals: their genetic strategies extend close kinship bonds across very large social spaces, enabling them to reap the fruits of specialization of labor. The only other path to sociality we find in nature involves high intelligence. It is surely a striking fact that there seem to be no non-eusocial social animals that aren’t perched atop steeply graded peaks in the landscape on which we plot cognitive sophistication[38]. Furthermore, the phenomenon is convergent[39]: all of the intelligent social animals – parrots, crows, rats, canines, pigs, elephants, hyraxes, raccoons, whales, monkeys, and apes – have more recent common ancestors with some asocial, unintelligent animals than they do with other social, intelligent ones. These two facts, taken together, suggest that sociality and intelligence mutually reinforce one another in evolutionary dynamics. This is a phenomenon that evolutionary game theory nicely explains: intelligence and sociality might mutually promote one another because non-parametric social problems[40] are computationally more demanding than the largely parametric ones confronted in the short run (i.e., among individual organisms) by the asocial[41]. It is also to evolutionary game theory, in conjunction with empirical historical ecology and genetics, that we must turn for explanations of why some populations were pushed up this hill and others weren’t. Sharks have been around for a very long time, but they haven’t been getting smarter and more social. Among the things that students who aim to study these phenomena need to do is work through the exercises in Gintis’s[42] textbook on evolutionary game theory; and this will not encourage them to model the origin of intelligence as emerging from Nash bargaining. So, here, advantage Gintis; but as I will argue later, there is no reason this should trouble Binmore.

As many authors – but especially Tomasello[43] and Sterelny[44] – have emphasized, humans do not just stand at one end of a continuum that slopes gently up the dimensions of intelligence and sociality, with mice and horses just up from the base and dolphins and gorillas standing only a few steps lower. Humans, and humans alone, have collectively accumulated fundamental behavioral changes, at a steadily accelerating rate, and have thereby ganged up in gigantic ensembles to take over the planet. The origin and the subsequent dynamics of human cultural evolution are distinct and additional phenomena that demand modeling and explanation over and above accounts of the genesis and stabilization of sociality. It is at this point that the issue of scale relativity I introduced in Section 1 becomes important.

As Sterelny[45] argues, one of the properties that crucially distinguishes people from what I called `BI’s’ is that the former, but not the latter, process information by way of decoupled representations. That is, they package information into salience classes that are not tightly linked to specific, stereotyped behavioral responses. This permits them to have beliefs about new kinds of objects, and desires over general states of affairs to which BI’s are insensitive. This is why we must model people as having different utility functions from their ancestral BI’s: the two sorts of utility functions will range over different salience classes of outcomes (though the classes may of course have elements in common). Furthermore, H. sapiens did not evolve the disposition to turn into people by growing neural wonder tissue[46]: we have the same basic control apparatus under the hood as our fellow mammals (though perhaps we have evolved a distinctive, dedicated module for handling language, as Sterelny concedes; see Ross[47] for an alternative discussion). What has instead happened is that we have collectively filled our environment with accumulated cultural artifacts, some physical but many virtual, that cue our behavioral responses and coordinate them around the abstract salience classes that are unique to our informational dynamics. Sterelny reserves the phrase `cultural evolution’ for this process, and distinguishes it from social evolution – that is, from the evolution of sociality itself. In moving from modeling social evolution to modeling cultural evolution one must represent a phase shift in the time scales. With respect to the longer-run evolutionary dynamics in our species, it may be just noise that Europeans domesticated pack animals centuries before Americans did, and that the former happened to get gunpowder and resistance to smallpox as well; but this is anything but mere noise when we are trying to explain patterns in cultural evolution[48].

Now let us come back to Binmore and Gintis with this perspective in mind. Both of them take up aspects of humans’ unique cultural dynamics, but in importantly different ways. Gintis recognizes that we are unusual creatures, and proposes to represent this fact by building prosocial preferences directly into our utility functions. Binmore seems at first glance to be denying distinctions like Sterelny’s when he defends the principle of methodological individualism in game theory on the basis of the fact that BI’s are the units of selection[49]; and this is the basis for all the difficulty between him and Gintis over whether people have self-regarding preferences. However, as I described in section 1, Binmore in fact does implicitly distinguish between social and cultural evolution. In fact, his doing so is the very core of his theory; it is Gintis who leaves the distinction between social and cultural evolution inside a black box. Let me explain.

Recall that for Binmore fairness norms evolve over what he calls the medium run – the time scale at which cultural institutions evolve – and condition people’s behavior in the short run. People are modeled as coordinating their behavior in repeated-game equilibria around these norms. Since people’s utility functions are to be inferred from their behavior in equilibrium on Binmore’s scheme, these norms – both before and after their moral content has been leached out by habituation – must enter into the utility functions in question. Their utility functions are thus not those of their ancestral BI’s – though features of these ancestors according to Binmore, explain many aspects of their utility functions that cultural evolution hasn’t yet had time to modify. Binmore’s theory is in part designed to show us how to model the processes whereby BI’s evolve into people; and the current people, to whom the normative content of the theory is addressed, are held to live partly under the shadow of the distant past. This invites some ambiguity in interpreting Binmore, into which, I hypothesize, Gintis stumbles. For Binmore, when we study long-run evolutionary dynamics using evolutionary games the relevant utility functions in contention are those of BI’s. BI’s are self-regarding, regulated by Hamilton’s rule. The behavioral dispositions of people – again, only in the long run – must be consistent with those of the BI’s with which they are coextensive. This is one sense – as we will see later, there is another – in which, for Binmore, people’s preferences are self-regarding. However, when Binmore’s evolutionary game theorist sets out to recommend institutions to people, the utility functions to which she is supposed to attend are those revealed in medium-run equilibria over the salience classes produced by the recent cultural evolution of the game theorist’s clients. The preferences entering into these utility functions will not be self-regarding in the sense that would justify Gintis’s critique, since local fairness norms will be built into them. Indeed, just so far as Binmore’s general theoretical approach is concerned, these preferences could be identical to the ones Gintis thinks are directly revealed in experiments with human subjects. For reasons we will take up later, Binmore just does not think that it is good methodology to try to recover the preferences so directly.

Gintis tells a slightly different story, one made clearer in a paper[50] that preceded his response to Binmore in this volume. According to him, the BI’s that were the ancestors of people evolved the capacity to program one another’s preferences. (This is Gintis’s phrase for social / cultural shaping of individuals’ preferences by other people. It collapses Sterelny’s distinction between socialization and enculturation.) Of course, Gintis requires that this evolution must have been consistent with Hamilton’s rule, and so models it. Programmed agents are said to have internalized norms. Some norms can be altruistic, in the sense that acting on the basis of them can systematically undermine fitness in the short run.[51] If enough altruists are present in a population, then altruism can be consistent with long-run fitness maximization. So, viewed statically, people can have altruistic utility functions. Of course, Gintis doesn’t deny that they have all sorts of standard mammalish goals and dispositions; but in his modeling framework it is just a brute fact which of these have been overthrown by social and cultural evolution and which have not. We are to discover the brute facts in question by putting experimental subjects into games and inferring their utility functions from whatever they tend to maximize in these games.

I said above that Gintis, unlike Binmore, shoves the evolution of people from BI’s into a black box. Based on the account above, this may be thought unfair. Doesn’t the account of the evolution of altruism as an exaptation on norm-internalization amount to Gintis’s way of opening the box? Well, only up to a point. Binmore starts to give us an account in game-theoretic terms of the relationship between the utility functions of BI’s on the long-run time-scale and the equilibrium-path dynamics of people’s utility functions on the medium-run (and, as we will see later, also on the short-run) time scales. (He only starts to do this because people are distinguished from BI’s by more than fairness norms, and these are all Binmore tells us about. Ross[52] takes the story further.) Showing, as Gintis[53] does, that personal altruism is statically consistent with dynamic equilibrium is not equivalent to modeling the evolution of people; it is a possibility proof that people can evolve. Here is where the difference emerges: Gintis, but not Binmore, hypothesizes a genetic modification in the hominid line, one that made them receptive to preference-programming. Evolutionary anthropologists are thus assigned the job of searching for the record of this modification. Binmore’s modeling approach need posit no such modification (though it also does not rule it out). In this, Binmore is consistent with Sterelny’s account. This is the crux of the disagreement. Gintis statically models people as we find them now, following a discontinuous evolutionary transition. Binmore models the feasibility space for human institutions given a continuous equilibrium path between their past games as BI’s and their medium-run cultural evolution as people.[54]

Now, what about the short run? As just described, Binmore is primarily interested, as our most clear-headed and rigorous contemporary political philosopher, in offering normative advice on medium run equilibria. On that time scale, short-run dispositions are noise. We should thus not infer medium-run maximization targets from what people appear to be maximizing in the short run. By contrast, Gintis, when he has his normative hat on[55] is interested in designing better mechanisms for us given our short-run dispositions just as we find them. Binmore in effect aims to show us how to stay on equilibrium paths while we culturally evolve better morality (by egalitarian lights); Gintis aims us to show us how to be better right now, if only we could establish superior institutions by collective action. It does not seem a cheap point here to note that Binmore calls himself a Whig, while Gintis identifies his political perspective with the socialist tradition.[56]

I think that the point I have just made explains a great deal. Even the most hard-headed scientific realist – and my head is comparatively pretty hard in this respect – should balk at the idea that there is a definite, objective fact of the matter over what people maximize at the short-run level of analysis. There are obviously all sorts of things they objectively do not maximize: their individual genetic fitness, for just one thing. (That is a target that fades away as we adjust the grain of analysis from long-run lineages and medium-run cultural representatives to individual people.[57]) But it is naïve to think that we’re going to discover that people behave so as to maximize (say) social status rather than their current ability to morally flatter themselves (by the lights of the current social contract) or their (hyperbolically discounted) medium-run freedom from anxiety, or various other ends that can be constructed under plausible descriptions. One thing you can learn from philosophy (of science, not from Kant) is that questions which have perfectly good answers when pressed at one grain of analysis do not continue to have good objective answers no matter how small you make the grain.

An analogy is helpful here. You can capture something useful about a person’s or an animal’s behavioral dispositions by stating its beliefs and preferences. Some of these attributions will be true and others false. So, for example, my dog might truly believe that he gets a walk after supper. He certainly doesn’t believe that Napoleon lost the battle of Waterloo. However, it doesn’t follow from this that one can keep usefully pressing the question of what the dog believes to ever finer analytical grains. Is it better to ascribe the belief that he gets to investigate the neighborhood smells after he eats than to ascribe the belief that we go for a walk after his supper? This question has no factual answer. Similarly, what we identify as maximization targets are, beyond a certain point, sensitive to pragmatic decisions about what descriptions to use.

To study people using economics, we must model them as maximizers – a point that Binmore doesn’t dispute as long we don’t insist they’re maximizing at every scale of analysis. Some models of maximization will be useless for all or most explanatory and predictive purposes; others will work well for particular questions within particular frames. Binmore and Gintis presume related but generally different frames for their questions, and so they find different maximization models useful. Here is the root of the irreducible pragmatic element in their dispute that I promised to identify.

Now I want to show how this element ramifies through all of the apparently empirical points of disagreement I enumerated at the end of Section 2.

4 Disputes about facts?

The diagnosis of the deep – but ultimately pragmatic – respect in which Binmore and Gintis differ was arrived at by starting to reflect on their disagreement (1), over the significance of punishment costs. Let us therefore begin the dissolution of the putatively empirical points of conflict by completing that reflection. Recall that, as Gintis argues, costly punishment in large n groups with private information must either imply low levels of equilibrium stability or implausibly high levels of second-order punishment. Now, if one is explaining the origins of sociality by means of game-theoretic models, the players in one’s games must be BI’s; their genetic priority will be reflected in granting them logical priority in the models. Since this is the problem that most early game-theoretic models of the logic of cooperation implicitly took themselves to be addressing, non-trivial punishment costs are the norm in such models. As noted early in the previous section, among BI’s punishment must be costly, because it can only be effective if it harms the punishee’s expected fitness. When such punishment is attempted, punishees who are BI’s will generally fight back. Furthermore, living in a society where punishment aims at reducing fitness implies that everyone bears costly risk of being punished. Unless the cooperative equilibrium in such a society is very stable, its expected welfare level must be significantly reduced by all the smiting that goes on. This will be especially true if non-discriminating trigger strategies, in which punishment of one implies punishment of all, are the only ones available.

However, on Binmore’s model fitness-reducing punishment need only go on at the long-run scale. It typically won’t be meted out to individual people in a form specifically targeted at them as individuals. In the long run, once people are enculturated, fitness-reducing punishment will generally occur in the context of inter-group competition. Individual people often won’t experience this subjectively as punishment at all, since people often don’t identify their own utility with their fitness, or with what they might think of as the interests of their population. It is, indeed, semantically odd to think of these dynamics as involving `punishment’ at all, since folk semantics prototypically associates the concept with short-run dynamics. Are present-day Europeans being `punished’ by low reproduction rates relative to Africans? This is an odd thing to say, but that fact is of no logical relevance in the context of formal modeling. If the behavior of Europeans is lowering their expected fitness, then in the technical sense they are being punished at that scale.[58]

This point, about what we find it natural to say, isn’t the important one. What matters is that in Binmore’s model of the evolution of norms, it is on the relationship between the short and medium runs that our analysis should focus. At this scale, the punishment of interest goes on in the short run. It occurs among people who are already enculturated, and needn’t bear the burden of bringing about socialization amongst BI’s. Once agents have already been socialized they can often be severely punished very cheaply. As Binmore continuously stresses, the most important punishment devices amongst people aren’t fines and prison terms, but frowns and sighs of disapproval. These are ineffective when deployed across ingroup / outgroup barriers, but among non-sociopaths have mortifying influence inside Binmore’s moral circles. Where stability is concerned, the devices have an additional attractive property along with being practically costless: they inflict no harm that cannot easily be undone, especially by the punisher. You can fill me with shame, and with terror of the gossip you might start, when your narrowed eyes tell me that you’ve caught me out in a bit of free riding. Then I ostentatiously cooperate with you – perhaps, but not necessarily, adding a small premium over and above the socially mandated level – and I get rewarded with a smile and a backslap. In normal cases, this amounts to complete withdrawal of the sanction and all its consequences. Furthermore, as a result of the whole episode we have exchanged some information, in both directions, about our respective power as bargainers, and about our mutual knowledge of conventions, and so about the exact location of equilibria in the immediate short-run game and in other games. Because your frown cost you so little, and did me no lasting harm, it is easy for us to achieve mutual gains from trade from an episode involving punishment; we don’t have to make up for its costs before taking our profits. In standard models in evolutionary game theory, where punishers simply pay costs and punishees suffer losses, these dynamics are not represented. I thus do not think that these models accurately represent games amongst culturally evolved people. By contrast, BI’s could care less about being frowned at.[59] Indeed, you can’t frown at a cow, except self-indulgently, because frowns are social actions and cows aren’t social.

None of this shows that stability is easy to achieve amongst enculturated agents. It merely shows that Binmore’s theory of moral-cultural equilibrium can’t be refuted by models that incorporate significant punishment costs and permanent losses due to punishment. These should be thought of as models of a different phenomenon on a different scale, namely, the evolution and maintenance of sociality.

Questions are also in order about stability itself. This refers to something quite precise in evolutionary game theory: the relative sizes of basins of attraction in a phase space. We lack, however, a stable metric for measuring this in applications to natural (as opposed to virtual) societies. Just how stable are the norms in large n human societies? These societies rise and fall. Should we think of them as lurching chaotically from crisis to crisis, or are what we call `crises’ just epiphenomenal storms on calm seas of deep coordination? Historians happily switch gestalts back and forth when making such decisions for the sake of plotting their narratives – here is the point about irreducible pragmatism again.

Binmore’s models are restricted to two-person communities – the smallest n in which the models can still make conceptual sense – and he is frank in acknowledging that scaling them to allow incorporation of coalition dynamics is an essential future task. Behavioral economists have mostly studied dynamics in small n groups, though often composed of members drawn from large n societies. Finding convergence on norms among such subjects indeed suggests something about stability. However, variation in responsiveness to norms is always observed. How much variation should be enough to lead us to think in terms of instability? What is our baseline here? It should not be all the logically possible vectors of strategies, since we know our subjects at least have plenty of common biological structure, and this will generate salience convergences that pare the set of possible vectors down from the infinite set allowed by the folk theorem. But what is the biologically feasible set of strategy vectors among people? Binmore, having no more idea than the rest of us, lets economic logic alone select the feasible set – for two-person communities. Gintis and other behavioral economists try to empirically discover the set by directly studying people as they play games.[60] Both approaches are useful for different purposes, as I will discuss below. But we are not in a position to say that a model of human norm dynamics must be incorrect unless it predicts a substantial degree of stability in large n societies. Can we rule out the hypothesis that such behavioral predictability as we find in such societies just results from the fact that people imitate each other, while the contents of norms otherwise drift about chaotically? I don’t think we can – though this doesn’t imply that we shouldn’t push stability hypotheses for all they’re worth and see what we get.

Once we recognize that costly punishment, as modeled in standard experimental games by allowing subjects to pay money to retaliate against free riders, isn’t perfectly analogous to the costless punishments people routinely inflict and withdraw in daily life, then the disagreement between Binmore and Gintis on point (4) from section 2 becomes less sharp. Many people clearly think that others are unduly tolerant of what they regard as moral laxity. Is this not most of the emotional content of so-called `conservatism’ in modern industrial societies? Why do people who aren’t conservative in the stereotypical sense often manifest respect for dogmatic moral beliefs of stereotypical conservatives when in their company? Is this not at least partly motivated, at least often, by the desire not to be clucked at? And do not most people, however avant-garde their explicit social morality, not act as conservatives do with respect to some marginalized attitudes they regard as louche? These phenomena are manifestations of everyday second-order punishment. It is of course interesting and important to find out that people are disinclined to pay cash to punish those who fail to punish free riders in public goods games, as Henrich et al[61] discover and as Gintis reiterates in his objections to Binmore. But this does not show that people don’t engage in second-order punishment in enforcing moral judgments where doing so has little cost. I have argued that the relevance of Binmore’s model doesn’t require any more than this.

Similarly, it is not at all obvious that Binmore’s model should predict that people’s behavior should always converge toward narrow selfishness in one-shot dictator and ultimatum games, as Gintis’s emphasis on disagreement point (2) suggests. In fairness to Gintis, Binmore certainly invites this emphasis by stressing the empirically observed tendency of people to learn to defect in one-shot PDs. As argued by Andreoni[62], however, one possible explanation for this is that the experimental setups in which these games have been studied are often ones in which people have no first-order punishment strategies available to them other than defection. Fehr and Gächter[63] interpret evidence from a public-goods game experiment in which they controlled this parameter as supporting Andreoni’s proposed effect. Notice, however, that Binmore’s model of fairness norms as medium-run Nash bargaining equilibria is only committed to the prediction of convergence toward selfishness given some specific assumption about the speed with which novel game environments shift culturally learned strategic dispositions. Nothing in Binmore’s model predicts any particular such coefficient; his short, medium and long runs are defined only relative to each other, not relative to any empirically derived measures of human cognitive plasticity. Exactly the same point applies to his disagreement (6) with Gintis.

Now, Gintis could appropriately reply here that this kind of response on Binmore’s behalf is ad hoc hypothesis shielding. Gintis’s thesis that H. sapiens underwent a discontinuous genetic transformation on the road to altruism predicts that the ratchet won’t be able to turn backwards, that people won’t `learn’ to revert to the utility functions of their ancestral BI’s except in the long run; whereas the defense of Binmore I have offered merely argues that he needn’t make any measurable prediction on this subject at all. We would indeed have to call `advantage Gintis’ here if Binmore had no good reason for incorporating expected reversion to selfishness into his model. However, as I will now argue in the concluding section of this paper, he does have such a reason, one that Gintis does not address.

5 Political philosophy and empirical induction

It is not Binmore’s primary purpose in Natural Justice, or in Game Theory and the Social Contact [64], to extend our empirically derived knowledge of human behavioral dispositions. As discussed earlier, Binmore’s is a project in political philosophy. He aims to show us that, in light of an empirically supported model of those behavioral dispositions, we do not confront a grim dichotomy between conservative refusal to risk stability for the sake of welfare improvements and egalitarianism, on the one hand, and a reckless utopian passion for leaping blindly into the dark, on the other hand. We can instead examine our current fairness norms, identify their equilibrium conditions, and then define forward equilibrium paths that dynamically promote both welfare efficiency and the relative well-being of the worse off. It is an intended virtue of this normative theory that it should persuade people to support its proposals for institutional reform along equilibrium paths if they endorse only limited concern for others. Indeed, they need only agree that it is vicious to resist Pareto improvements. The rich and powerful, Binmore seeks to show, can venture forward from the status quo without having to sacrifice expected utility at any point; this is the property that makes Binmore’s reform path an equilibrium path. (Note carefully that, on Binmore’s scheme, the rich will have to surrender some wealth. They are incentivized to do this in exchange for gains in security from appropriation, which is why they lose no utility. However, as I noted in the opening section of the paper, this incentivizes them to try to keep the poor from learning about their own bargaining power, if possible. This explains the sound intuition that conservatism crucially rests on suppression of truth. If Binmore’s analysis were common knowledge, then the self-interested rich should sign on to the new social contract. So the self-interested rich should not recommend Binmore’s work to the poor or to the altruistic rich.)

It should be noted that this point addresses the one critical objection that Paul Seabright[65], after endorsing Gintis’s criticisms of Binmore that I have been discussing, adds to the arsenal. Seabright wonders what the functional point of a theory of justice could be on Binmore’s account, if levels of second-order punishment are high enough to make cooperation an equilibrium strategy for self-interested agents. The answer lies in Binmore’s scale-relativity and sensitivity to dynamic relations among the scales. What the powerful gain in return for the concessions they make is a stabilizing social contract in which the less powerful view the equilibrium level of inequality as just and so do not challenge it. For Binmore, moral intuitions are inherently conservative – a view that itself seems to line nicely with everyday intuitive judgments about the uses of morality in politics. Fortunately, since the status quo is always moving, and since medium-run adjustment lags short-run adjustment, the theory of justice is always open to improvement along the equilibrium path. This is the value of political philosophy as Binmore practices it.

Binmore’s strategy here reflects two excellent principles to which all political philosophers should swear oaths. The first is Hume’s principle. It tells us that wherever we can accomplish our ends using either, on the one hand, institutions that can be undone by the morally unscrupulous, or, on the other hand, institutions that are proof against rational knavery – that is, narrow selfishness – then we should choose the latter. The second principle is that useful political critique does not consist in condemnation of the status quo from godlike perspectives that suggest that the critic is not herself a player in the social game of life. Reformers should instead propose specific bargains to specific agents who hold preferences that differ from their own. It accomplishes nothing when Kantians preach to Kantian choirs, and Marxists to Marxist ones, and libertarians to libertarian ones. Most people are not Kantians or Marxists or libertarians and are never going to be. The political philosopher, Binmore often stresses, should address people as they actually are, if institutional reform is truly what she cares about.

There is some tension between these two principles, because – just as Gintis emphasizes – though people are not disinterested philosopher-kings, most of them are also not Hume’s knaves. Gintis and other behavioral economists aim to empirically discover just where most people actually stand, right now, between these poles. This is of course a useful thing to find out as much as we can about. But perhaps it is naïve to think that political reforms should be aimed at people exactly as we think they are now. Part of the point here is that there is a grain of analysis below which there is no way they exactly are that can be captured in the formalism of utility theory. But the relevant aspect of the point for immediate purposes is that one thing we know about how they are is that they are quite plastic. Perhaps the very fact of learning some evolutionary game theory will make them less selfish, as they learn about feasible cooperative equilibria they hadn’t previously recognized. But it is equally possible, for all we know, that as they become sophisticated in their theoretical knowledge (as opposed to their knowledge of the laboratory incentives) they might become more selfish.[66]

This suggests an interesting way of deciding whether to be a whig or a socialist. Suppose you think that there are enough people out there who can learn to be effectively, strategically, self-interested that if you don’t make your institutions proof against them, they’ll tend to find and exploit any loopholes allowed in your reforms. I take it that most astute behavioral scientists, including Gintis, will find this supposition reasonable. In light of it, here is the $64-million question for political meta-ethics: Are the expected welfare costs of building in knave-proofing devices against the loopholes higher or lower than the expected welfare costs of (possibly) constricting the set of feasible social contracts by closing the loopholes? Binmore’s whiggery consists in a bet that the opportunity cost of loopholes is higher than the opportunity cost of the feasible-set restrictions. This bet is in turn based on his hypothesis that our evolutionary development from BI’s to enculturated people was continuous rather than discontinuous, since this in turn predicts that people will find it easier to learn to be effective knaves.

I find nothing in Gintis’s writings to suggest that he is making the opposite bet. That is, I don’t impute any derivational link between his (broad, imputed – see note 56) socialism and his belief in an evolutionary discontinuity. I hypothesize, rather, that he takes it as axiomatic that behavioral scientists should just aim to find the descriptive facts; and then what he takes to be the descriptive facts happen to look like good news for socialists. So I am not accusing him of indulging in a non sequitur based on wishful thinking. I instead intend two critical points that are a bit more subtle.

First, we need to distinguish between two kinds of empirical claims Gintis makes, which have quite different levels of evidential weight behind them. The data indeed show that people might not be as quick to learn their way out of socially programmed utility maximization as Binmore believes. On one-shot PDs Binmore has solid data on his side, but the Fehr and Gächter results on public-goods experiments show that the Andreoni effect is indeed a possible explanation of these data. However, Gintis’s other empirical claim against Binmore, his hypothesized discontinuity on the road to altruism, is just a case of inference to (what Gintis takes to be) the best explanation. I think that Sterelny’s alternative such inference, which supports Binmore, handles more behavioral and evolutionary considerations than Gintis’s. Second, even if Gintis were right about the discontinuity hypothesis, this would at best show that we need to do further modeling work to test between the whig and the socialist bets where normative theory is concerned. Here, Binmore surely has a great deal of everyday observation on his side. People cheat and chisel one another quite a lot, in full awareness of the fact that social morality preaches against their doing so; and overconfidence in human prosociality helped to kill an enormous number of people during the past century. This is not to deny – and Binmore does not deny – that conservatism also has a great deal of blood on its record. But whiggery is anti-conservative too. Whiggery is nervous egalitarianism.

Now, the claim `people cheat and chisel quite a lot’ hardly has scientific content. How much is a `a lot’? Just how harmful is the cheating people get away with in modern large n societies, relative to the opportunity cost of being cautious in our reform plans? We should use game theory to try to give the claim some content. For any given model of the status quo, we can use bargaining models of Binmore’s sort to ask: how much bigger is the feasible set of social contracts if we plug in one of Gintis’s favored utility functions rather than Binmore’s? But how much more sensitive to errors is the stability of the new equilibrium path? The fact that Binmore’s models abstract away from coalitional dynamics is relevant here. In multi-agent societies, effective exploitation of social contract loopholes over the long run requires coalitions to coordinate on knavery, something not necessary in 2-person bargaining games. But how difficult is such coordination in the medium run given near-zero punishment costs? – this is the actual question Binmore’s specific normative inquiry poses. As far as I know, the work we’d need to do to answer this question remains outstanding.

We can now shed some light on dispute point (5) from section 2, the only one that still remains undiscussed. If the folk share (to some extent) Binmore’s concern to make their institutions knave-proof, this can explain why they’re disinclined to be generous to the poor except where they’re confident that the poor are innocent victims. If people embrace egalitarian fairness norms, as both Gintis and Binmore agree they do, then these very norms are open to exploitation by free riders. If I know that you will take care of me if I shirk, then I have some incentive to shirk – especially if I discount hyperbolically and thus don’t give rational weight to the loss of self-esteem my own internalized fairness norms might produce in me after a few months on the dole. Of course, this reasoning is the oldest trick in the conservative’s book. It is morally urgent for us to recognize, following the brilliant work of von Parijs[67], that the free riding represented by shirking has lower welfare costs, given any reasonable model of actual human preferences, than the Calvinistic festishization of work for work’s sake that turns most people’s lives into red queen games. However, the important point at the moment is that the effectiveness of the conservative’s old saw is explicable. Now, is Gintis right that popular fear of incentivized shirking only makes sense given preferences that aren’t self-regarding in Binmore’s sense? Why should a self-regarding agent punish potential shirkers in order to safeguard the social contract? Doing so looks like altruism.

But wait: telling pollsters that one opposes welfare for shirkers, or voting for politicians who promise to find the shirkers and take their benefits away (for tax savings) hardly constitutes costly punishment behavior. Do typical conservative voters give up their leisure time to troll the bars for people who are frivolously spending their welfare cheques and turning down jobs? Once a welfare state of sorts is in place, does anyone pay costs to punish those perceived to be shirkers? Blocking their access to the social insurance pool probably, in general, saves the institutions that administer the punishment more resources than it costs them. Once again, then, and for the same reason as has been invoked before, I don’t see that Gintis’s model of people here predicts data that Binmore’s model counter-predicts.

As I noted previously, disagreement (3) is the fundamental element of the set, the one that drives the others given special assumptions about punishment costs and other empirical variables. I said early on that two methodological divergences underlie (3). The first of these has just been discussed. Binmore’s hypothesis about the speed of short-run strategy adjustment is his way of encoding Hume’s principle and the no-philosopher-kings principle. This hypothesis must be made consistent, for modeling purposes, with the representation of various selfish strategies as being medium-run equilibrium strategies under a variety of circumstances, since medium-run equilibria are the reference points Binmore takes for the revelation of utility functions. It then just follows that some short-run cooperative behavior Gintis will use to derive utility functions will be treated by Binmore as noise. But Binmore has an additional methodological motivation for his approach to utility assignment, one suggested in the remark I quoted earlier that “I would be discouraged to find that each new experiment seems to require subjects to have a brand new utility function, but behavioralists are made of sterner stuff than me!”[68]

Binmore[69], describing the relationship between the version of homo economicus in his model and Homo sapiens, says of the latter “Much empirical evidence can be marshaled in favor of the proposition that homo sapiens has ethical propensities built into his software or hardware. He votes. He contributes to charity and to public television stations. He tips in restaurants that he does not anticipate visiting again. In many countries, he donates blood without monetary reward. Sometimes, he will risk his life in an attempt to save a total stranger. Sometimes, he will sacrifice his life altogether in the name of some abstract cause or principle. The evidence, both factual and theoretical, for the proposition that homo sapiens shares some of the virtues of homo ethicus seems overwhelming. However … one does not necessarily need to step outside the homo economicus paradigm to accommodate most of the evidence.” Everything up to the final sentence could be Gintis speaking. Let us therefore focus on that final sentence, and assume that its choice of words is not casual. The evidence is to be accommodated by our modeling formalism, not predicted by it. Binmore often makes remarks that suggest the main point I have been emphasizing in this essay: use of a representational formalism, such as utility theory, always reflects a pragmatic choice on an analyst’s part. Such choices ought to be constrained by the structure of the world to be modeled, but they are never forced by it.

Logical positivists – or caricatures of them – have been getting terrible press for several decades now. As a result, many people have forgotten some useful distinctions they made. One of these was between the construction of a formal theory and the empirical interpretation of that formalism. The philosopher W.V. Quine was merely reiterating a point familiar to his great mentor, Carnap, when he argued that, given some formalism, decisions about which scheme of interpretation to use in applying it to particular empirical phenomena are irreducibly pragmatic.[70] It is easy to forget this, because the pragmatic decisions in question are often no-brainers as a consequence of the strong constraints imposed by the history of scientific practice; one doesn’t have to interpret the basic objects of reference in General Relativity theory as fields, but it would be perverse to ask everyone to invest in an alternative. The pragmatist’s point is practically significant, however, when we engage in theoretical unification, something both Binmore and Gintis explicitly claim to be doing.

I suggest that Gintis, in his criticism of Binmore, is forgetful of the pragmatist’s point. He appears to assume that the obviously best way of mapping utility theory onto human behavioral phenomena is whatever way yields the most direct and straightforward description of the totality of current observed behavior across the empirically available range of situations. One pragmatic consideration that he thereby forgets is scale relativity: it is the standard, not the deviant, case in science that the same phenomena get sorted into different salience classes at different scales of generalization: think of general relativity theory versus quantum mechanics and molecular genetics versus population genetics. Now, we should not pester scientists about this sort of thing for nothing but philosophical reasons; scale relativity should be set aside until and unless some scientific motivation makes it relevant to measurement and to interpretation of some formalism. But Binmore does have such a motivation: the relationships between short-run, medium-run and long-run equilibrium dynamics are the principal phenomena he has set out to model. He knows that he could represent human behavior in game-theoretic terms by deriving utility functions from short-run equilibrium assumptions; but then, if still wanted to develop his normative theory within the constraints of Hume’s principle, he’d have to go outside the standard economic formalism and impose it exogenously.

Could Gintis make a convincing pragmatic case that that is just what Binmore should do? Such a case would have to depend on an argument that Gintis’s pragmatic preference leaves fewer empirical data underdetermined than Binmore’s does. Gintis’s critical focus on empirical disagreements rather than methodological ones suggests that he implicitly recognizes this, and aims to rise to just that challenge. However, this chooses one side of a trade-off. Once Binmore references utility functions to medium-run equilibrium strategies, he has no further discretionary freedom with respect to modeling the short run: departures from Nash equilibria there have to be treated as noise of one sort or another. By contrast, Gintis is free to hypothesize – and then empirically test, of course – alternative models of short-run maximization. But the irreducibly pragmatic element can’t be made to go away here. There is no unique, empirically measurable salience class from which to distinguish a finite set of alternative micromotivational targets. Again, there are constraints; people can’t try to maximize over salience classes they can’t informationally distinguish, and now that we have neuroeconomics to tell us about this we can expect it to progressively rule out various hypotheses that are presently open. As Camerer[71] suggests, that is good news for the research program he shares with Gintis, and a reason their `stuff’ doesn’t have to be as Quixotically stern as Binmore implies. He should not try to talk them out of their approach on a priori methodological grounds. (Beware, indeed, of the philosopher’s impulse, I preach to the choir.) But I have tried to show that he has excellent reasons for locating his own pragmatic modeling freedom at a different point. We have still to see how this works out when coalitional dynamics are explicitly introduced into the framework. So far, however, my personal verdict is that Binmore’s work is the most promising development in political philosophy we have had in a very long time.