Why We Punish Strangers: The Evolution of Human Cooperation

Cooperating with strangers is the trait no other primate scales up

A person in a modern city routinely cooperates, in a single day, with hundreds of genetically unrelated strangers she will never meet again and who could never retaliate against her if she cheated: the driver who yields, the cashier who gives correct change, the stranger who returns a dropped wallet. None of that is explained by shared genes, and most of it is not explained by the prospect of meeting the same person twice. It is also not unique to markets or money. Across foraging bands, horticultural villages and herding camps that have never seen a bank, people share meat with non-relatives, contribute labor to projects whose benefits they cannot exclude anyone from, and — the fact this article is built around — will spend their own resources to punish someone who defected against a third party, with no prospect of getting the cost back.

Set against the rest of the animal kingdom, this is not a matter of degree. Melis and Semmann’s review of the comparative literature states the anomaly plainly: cooperation is a widespread phenomenon in nature, but “human cooperation exceeds that of all other species with regard to the scale and range of cooperative activities,” and cooperation between unrelated individuals extends in humans even to complete strangers who are not likely to meet again in the future ^[10]. Chimpanzees, humans’ closest living relatives, are by their own account “one of the best candidate species” among non-humans for the kind of contingency-based reciprocity this article traces below — and even there, the empirical evidence for it “remains scarce” ^[10]. Intraspecies cooperation between unrelated individuals outside humans generally is, on the same review’s account, comparatively uncommon ^[10]. Whatever produced the human pattern, it did not simply amplify a mechanism already doing serious work in a chimpanzee troop.

This article is not a tour of settled science. It follows the theoretical floor that kin selection and simple reciprocity put under cooperation, the experimental program built to measure what happens above that floor, and — because this is a genuinely contested field — the sharpest critics of that program on their own terms, not as a footnote. It ends with three live candidates for the mechanism that does the remaining work, presented as a real disagreement rather than a resolved one, because that is the state the evidence actually supports.

Kinship explains the family, not the marketplace

The baseline case is the one evolutionary biology solved first, and it solved it exactly. W. D. Hamilton’s 1964 papers on the genetical evolution of social behaviour gave altruism toward kin a precise condition under which natural selection favors it ^[1]:

Here $r$ is the coefficient of relatedness between the actor and the recipient, $b$ is the fitness benefit the recipient gains, and $c$ is the fitness cost the actor pays. A gene that disposes its carrier to help a relative will spread whenever the benefit to the recipient, discounted by how likely the recipient is to carry a copy of the same gene, exceeds the cost to the helper. As Robert Trivers summarized it a few years later, opening his own paper on the harder problem of altruism between non-relatives, “Hamilton (1964) has demonstrated that degree of relationship is an important parameter in predicting how selection will operate” on behavior that looks altruistic ^[2]. Under Hamilton’s rule, a parent who works to feed a child, or a sibling who takes a real risk for a full sibling, is not a puzzle. The genes involved are, in the relevant sense, helping copies of themselves.

What the rule does not do is scale. Relatedness falls by half with each step of genealogical distance — a full sibling carries $r = 0.5$, a first cousin $r = 0.125$, and by the time two people are unrelated in any traceable genealogical sense, $r$ is effectively zero. Set $r \approx 0$ into Hamilton’s inequality and the right-hand side has to do all the work: either the benefit to the recipient must be extraordinary or the cost to the actor negligible, for almost every real transaction between strangers. That condition does hold for some human behavior — warning a stranger about to step in front of a car costs almost nothing and saves a great deal — but it manifestly does not hold for the transactions that build cities and nations: financing a stranger’s business, paying taxes into a fund one may never draw on, or absorbing the cost of punishing someone who cheated a person one has never met. Kin selection is not wrong; it is answering a narrower question than the one a market or a state actually poses. Any account of large-scale human cooperation has to start by admitting that this particular, well-confirmed mechanism runs out well before it gets there.

Reciprocity explains the neighbor who stays, not the stranger who leaves

The next mechanism in the textbook sequence removes the kinship requirement but keeps a different one: a future. Trivers defined reciprocal altruism as “behavior that benefits another organism, not closely related, while being apparently detrimental to the organism performing the behavior,” where the apparent cost is repaid later by the recipient reciprocating in turn ^[2]. The logic does not require shared genes; it requires that the same two individuals interact again, and that the returned favor be worth more to the original altruist than the favor cost.

Trivers was explicit about what has to be true of a species, or a social setting, for this to work, laying out three conditions that make reciprocal altruism selectable: a long lifetime, so that any two individuals are likely to encounter many situations calling for it; a low rate of dispersal, so that individuals keep meeting the same small set of others rather than scattering; and a high degree of mutual dependence, which keeps individuals near each other in the first place ^[2]. Robert Axelrod and Hamilton gave this intuition a formal, evolutionary backbone a decade later in Science, showing that a simple conditional strategy — cooperate first, then do whatever the partner did last time, the strategy known as tit-for-tat — can invade a population of unconditional defectors and, once established, resist invasion by other strategies, provided the “shadow of the future” is long enough: the probability of meeting the same partner again has to be high relative to the one-time temptation to cheat ^[3]. Cooperation, on this account, does not need trust or morality to get started; it needs enough repetition that betrayal costs more than it gains.

That repetition requirement is exactly what a modern anonymous transaction lacks, and it is why reciprocity theory cannot by itself explain a driver yielding to a car she will never see again. Reciprocity scales to a village, a workplace, or a long-standing trading partnership, where reputational information circulates and the same names keep reappearing. It does not scale to a one-shot encounter with a genuine stranger, because there is no future interaction to make defection costly and no channel for information about the encounter to reach anyone who matters. Something else has to be doing the work in exactly the cases that define a large, anonymous, non-kin society — and the leading candidate, for the last three decades of research on this question, has been the willingness to punish.

The laboratory found strangers who pay to punish other strangers

Ernst Fehr and Simon Gächter built an experiment specifically to isolate that willingness from every mechanism already covered above. Two hundred forty university students played a public-goods game in groups of four, with fresh, anonymous group membership every round and instructions telling subjects they would never interact with the same person twice — a design built to rule out both direct reciprocity and reputation formation ^[4]. Each member received an endowment of twenty monetary units per period and chose how much of it, from zero to twenty, to invest in a shared project; every unit invested returned 0.4 units privately to the investor but 1.6 units split across the whole group, so the group as a whole did best if everyone invested everything, while any single member did individually best by investing nothing and free-riding on the others ^[4].

After investment decisions were revealed, subjects in the punishment condition could spend points to reduce other members’ earnings: each point cost the punisher one monetary unit and cost its target three ^[4]. This is the operational definition of altruistic punishment in this literature — a sanction that reduces the punisher’s own payoff, benefits no one materially, and is aimed at someone the punisher will never interact with again. It happened constantly. Across the ten experimental sessions, subjects punished other group members 1,270 times in total; 84.3 percent of subjects punished at least once, and the punishment fell overwhelmingly — 74.2 percent of all sanctioning acts — on below-average contributors, delivered by above-average contributors ^[4].

The consequence for cooperation was large and directional. Ninety-four percent of subjects individually invested more, on average, when the punishment option was available than when it was not, and the two conditions produced opposite trends over the six rounds of play: in the punishment condition, average investment rose over time, and by the final period 38.9 percent of subjects contributed their entire endowment and 77.8 percent contributed fifteen units or more; without punishment, investment fell over time, and by the final period 58.9 percent of subjects contributed nothing at all and 75.6 percent contributed five units or less ^[4]. In the paper’s own framing, cooperation flourishes if altruistic punishment is possible, and breaks down if it is ruled out ^[4]. When Fehr and Gächter probed the emotional mechanism directly, presenting subjects with hypothetical free-riding scenarios, 47 percent rated their anger at a low-contributing group member at the maximum end of a seven-point scale, and another 37 percent rated it one step below that — a pattern the authors read as evidence that negative emotion, not calculated future benefit, is the proximate trigger for the punishment ^[4]. This is one of the most replicated results in experimental economics, and it is the empirical anchor for the entire “strong reciprocity” research program discussed below.

Fifteen societies show the offer tracking a market, not a universal instinct

A separate and equally important question is whether the behavior Fehr and Gächter measured in Zurich undergraduates is a human universal or a product of the specific population tested. Joseph Henrich and a team of a dozen field researchers set out to answer that directly, running a common protocol — principally the ultimatum game, in which one player proposes a division of a stake and a second can accept or reject it, rejecting anything meaning both get nothing — across fifteen small-scale societies on five continents, spanning three foraging groups, six that practice slash-and-burn horticulture, four nomadic herding groups and three sedentary small-scale agriculturalist societies ^[5].

The first and least surprising finding was that the textbook prediction — a rational, self-interested proposer offers close to nothing, and a rational responder accepts any positive amount — was not supported in a single society studied ^[5]. The more consequential finding was how much the societies differed from each other, far more than had ever been observed comparing university students across countries: mean ultimatum offers in the sample ranged from 26 percent of the stake to 58 percent ^[5]. The low end was the Machiguenga of Peruvian Amazonia, described in the paper as “almost entirely economically independent” at the household level, with little productive activity extending beyond the family; the high end was the Lamalera of Indonesia, whale-hunters who go to sea in large crews of a dozen or more and depend on developed norms for splitting a joint catch ^[5]. Rejection behavior varied just as sharply: the Hadza of Tanzania rejected 24 percent of all offers and 43 percent of offers at or below 20 percent of the stake, while the Achuar, Aché and Tsimané accepted essentially every offer proposed to them ^[5].

The authors then tested what predicted a society’s mean offer, using two variables constructed from ethnographic description rather than from the experiment itself: how much of daily economic life depends on cooperative production beyond the household (“payoffs to cooperation”), and how much people rely on market exchange rather than household self-sufficiency (“market integration”). Both measures carried large, statistically significant, positive regression coefficients of similar size — normalized coefficients around 0.3 each — and together explained 68 percent of the variance in group mean offers; individual-level variables measured within each society, including age, sex, relative wealth, village population and experimenter identity, were not significant predictors of an individual’s offer ^[5]. Put plainly: what predicted how generously a person split a windfall with a stranger was not who that person was, but how integrated into cooperative production and market exchange the society around them was. An expanded treatment of the same fifteen-society dataset, with open peer commentary from across the field, appeared four years later in Behavioral and Brain Sciences ^[6]. Fairness toward strangers, on this evidence, tracks the economic structure a person lives inside rather than expressing a fixed, universal instinct at a fixed strength.

Most of what is known about fairness comes from one unusual slice of humanity

That result carries an uncomfortable implication for every ultimatum-game study run before 2001 on university students, and Henrich returned to the point directly nine years later with two co-authors, this time auditing the sampling frame of the behavioral sciences as a whole. Their analysis of the top journals across six psychology subdisciplines from 2003 to 2007 found that 68 percent of subjects came from the United States, and a full 96 percent came from Western, industrialized countries — nations that hold only 12 percent of the world’s population ^[7]. The skew traced back to the authors publishing the studies: 73 percent of first authors were at American universities, and 99 percent were at universities in Western countries ^[7]. Even within the West, the typical subject was unrepresentative of the West itself — in the Journal of Personality and Social Psychology, the flagship social-psychology journal, 67 percent of American samples and 80 percent of samples from other countries were composed solely of undergraduates enrolled in psychology courses ^[7].

The authors coined the acronym WEIRD — Western, Educated, Industrialized, Rich, Democratic — for this narrow, thin, and by their own account rather peculiar slice of the species, and their review explicitly lists fairness and cooperation among the domains where the comparative evidence shows WEIRD subjects behaving as frequent outliers rather than as typical representatives of the species ^[7]. This is not an outside critique of the punishment literature; it is the same research program auditing itself, since Henrich co-authored both the fifteen-society ultimatum project and the WEIRD paper. Read together, the two results say something sharper than “people differ”: the modal fifty-fifty ultimatum split long treated in economics as evidence of an innate human fairness instinct was substantially a fact about market-integrated, educated undergraduates, not a species-wide baseline — and the true range of the behavior was invisible until researchers deliberately left the sampling frame that had produced almost all of the prior evidence.

The ethnographic record is strangely quiet about costly punishment

If altruistic punishment is a psychological disposition shaped by natural selection because it once sustained cooperation in the small, face-to-face societies humans evolved in, ancestral small-scale societies should show it operating outside any laboratory. This is exactly where Francesco Guala’s 2012 target article in Behavioral and Brain Sciences delivers the most serious challenge in this literature, and it deserves to be stated as sharply as he states it ^[8].

Guala’s opening move is a distinction he argues the field has been running together: a narrow reading of punishment experiments, under which they are simply useful devices for measuring psychological dispositions under controlled conditions, is uncontroversial and is in fact how some of the field’s own leading practitioners describe their purpose — Camerer and Fehr write that their aim is “to describe a menu of experimental games that are useful for measuring aspects of social norms and preferences,” and Fehr and Schmidt separately emphasize that the games’ value lies in their simplicity, which “makes inferences about subjects’ motives more convincing” ^[8]. A wide reading — that these lab mechanisms are the same mechanisms that sustain cooperation in the world outside the lab — is the load-bearing claim the strong-reciprocity program actually needs, and it is a claim some of the same researchers do make explicitly: Fehr and Gächter themselves write, in Guala’s quotation of their own words, that “in our view punishment of free-riding also plays an important role in real life” ^[8]. Guala’s argument is that the narrow claim is well supported and the wide one currently is not.

The evidence he marshals is ethnographic, and it cuts against casual assumptions in both directions. Bowles and Gintis assert that “studies of contemporary hunter-gatherers and other evidence suggest that altruistic punishment may have been common in mobile foraging bands during the first 100,000 years or so of the existence of modern humans,” but Guala traces their own citation back to a source (Boehm 1999) that does not, on inspection, endorse a costly material-punishment account of foraging-band sociality ^[8]; Richerson and Boyd’s claim that “in small-scale societies, considerable ethnographic evidence suggests that moral norms are enforced by punishment” traces, similarly, to references that support punishment in general rather than costly punishment specifically ^[8]. The most systematic field attempt to settle the question is Polly Wiessner’s coding of sanctioning episodes among Ju/'hoansi foragers in Botswana: across a sample of 171 recorded episodes, only 8 percent carried any negative material consequence for the person doing the punishing, and Guala summarizes the overall pattern as one in which “the statistical incidence of material cost for the punishers is close or equal to zero” ^[8]. Colin Turnbull’s classic ethnography of Mbuti pygmy hunting bands supplies the sharpest illustrative case: a hunter named Cephu who repeatedly positioned himself to intercept game before it reached the group’s nets was sanctioned not by any one individual absorbing a cost, but by the whole camp’s coordinated ridicule and criticism, escalating to a meeting that stripped him of his illicit catch — after which reconciliation followed within the same day, and no one who took part in the sanctioning individually lost anything ^[8]. Guala’s broader synthesis across the historical and anthropological record is that real social dilemmas are far more often solved by institutions that coordinate sanctioning and keep its cost low and shared than by the uncoordinated, individually costly punishment the laboratory reliably produces — and that above roughly a one-to-three cost-to-impact ratio, the same ratio Fehr and Gächter used, further increases in sanction severity stop raising cooperation and start risking destructive feuds and counter-punishment instead ^[8].

None of this makes the laboratory result wrong. Guala is explicit that costly punishment in the lab is “robust to replication, a real experimental phenomenon” whatever it turns out to explain outside it ^[8], and the strong-reciprocity camp’s own strongest reply — that skepticism about external validity, absent new field data, is an argument that could stop any experimental inference rather than a specific refutation — is a fair point he concedes rather than dismisses ^[8]. What the ethnographic gap does establish is that the wide claim — that this laboratory disposition is what actually keeps cooperation running among strangers in the world — is, for now, an inference well ahead of its field evidence.

Even the word costly is doing double duty in this literature

Part of why this dispute persists is that the two camps are not always disagreeing about a fact; sometimes they are using the same word to make different kinds of claims. Guala separates an absolute cost of punishing — a fee paid in material terms, regardless of context — from a relative cost — the difference between what a punisher earns and what a non-punisher in the same position would have earned ^[8]. A punishment can be costly in the second sense while being close to costless in the first, exactly as Wiessner’s and Turnbull’s field material shows: symbolic sanctions like gossip, ridicule and ostracism impose almost no material burden on the person administering them, yet they remain disadvantageous relative to simply looking away, which is enough to generate the theoretical puzzle strong-reciprocity theory was built to solve, without requiring that anyone actually hand over real resources ^[8]. Much of the literature on both sides, in Guala’s reading, conflates these two senses when a study is cited as evidence for “costly punishment” in general.

The deeper disagreement, though, is real and unresolved rather than merely semantic. Strong-reciprocity theorists — the label attaches most closely to Samuel Bowles, Herbert Gintis, Fehr, Robert Boyd and Peter Richerson — argue that a disposition to cooperate and to punish defectors even at material cost, in one-shot and anonymous settings, is what lets cooperation survive precisely where the folk-theorem mechanisms of repeated-game reciprocity cannot reach: interactions with strangers, heavily discounted futures, and settings where information about past behavior does not circulate ^[8]. That is a genuine theoretical advantage weak-reciprocity accounts cannot claim even in principle, since they require exactly the long, well-monitored relationships that a modern anonymous transaction lacks. What remains open is whether the mechanism that gives strong reciprocity that theoretical reach is the same mechanism actually operating in the societies where large-scale stranger cooperation is observed, or whether — as the ethnographic record above suggests — those societies instead built cheaper, coordinated substitutes for it. Both claims can be true of different settings at once, which is exactly why this dispute has not, and may not soon, resolve into a single winner.

Three candidates compete to carry the load reciprocity cannot

If kin selection is too narrow, dyadic reciprocity does not scale to strangers, and the wide reading of costly-punishment experiments currently outruns its field evidence, something still has to explain why large anonymous human societies function as well as they demonstrably do. Three families of account are active in the literature, they are not simple restatements of each other, and no published synthesis currently assigns them agreed relative weights.

The first is cultural group selection, formalized by Robert Boyd, Herbert Gintis, Samuel Bowles and Peter Richerson. Their starting problem is that punishing free-riders is itself a public good — everyone in a group benefits if free-riders are deterred, but only the punisher pays the cost of deterring them, so a population of willing punishers should in principle be invaded by cooperators who never bother to punish, a “second-order” free-rider problem sitting on top of the first ^[9]. Their proposed way out rests on an asymmetry between the two roles. A plain cooperator who does not punish pays a fixed cost, call it $c$, simply by cooperating, regardless of how many other group members defect. A punisher, by contrast, only pays a punishment cost when there is actually a defector present to punish, so as defectors become rare within a group, the punisher’s fitness disadvantage relative to a non-punishing cooperator shrinks toward zero even while the plain cooperator’s disadvantage relative to a defector does not. Writing $x$ for the frequency of cooperators, $y$ for the frequency of punishers, and $k$ for the cost of punishing a single defector, the punisher’s expected fitness penalty is approximately

which falls as the residual pool of defectors, $1 - x - y$, shrinks. Once a group happens to reach a state where

punishing costs less at the margin than plain, unpunishing cooperation does, and ordinary within-group selection stops working against the punishing type in that group. If groups that reach this state also out-compete or out-reproduce groups that never do — amplified, the authors argue, by cultural transmission mechanisms such as conformist imitation of locally common strategies, which are not available to purely genetic evolution — group-level selection can carry punishment, and the large-scale cooperation it sustains, to frequencies individual selection alone would never reach ^[9]. This is an elegant resolution of the second-order free-rider problem on its own terms, and it is also exactly the kind of multi-level selection argument that a long tradition in evolutionary biology remains skeptical can operate strongly enough in real human populations, given realistic rates of migration and group extinction, to do the work the theory assigns it — a live theoretical objection the model’s internal consistency does not by itself settle.

The second candidate relocates the explanatory weight from evolved psychology to institutional design. Elinor Ostrom’s Nobel-winning body of work argued against both of the standard textbook solutions to commons dilemmas — pure privatization and pure centralized state control — on the grounds that long-surviving, self-governing communities converge instead on a recognizable bundle of institutional features that make sanctioning cheap, legitimate and coordinated rather than individually costly ^[11]. A concrete historical case Guala uses to make exactly this point is the Carte di Regola, communal-pasture charters used by villages in the Trentino region of northern Italy from around 1200 until Napoleon abolished them in 1800: they defined membership, funded designated guards to monitor compliance, and imposed graduated fines that escalated with repeated violations and fell mostly on outside trespassers rather than fellow villagers ^[8]. As Ostrom herself put it, summarizing decades of comparative case work: “In all known self-organized resource governance regimes that have survived for multiple generations, participants invest resources in monitoring and sanctioning the actions of each other so as to reduce the probability of free riding” ^[8]. On this reading, what looks in the laboratory like an individual, evolved willingness to absorb a personal cost is, in the societies that actually sustain large-scale cooperation for centuries, mostly engineered away into a shared institutional cost before it ever reaches an individual punisher.

The third candidate keeps the explanation inside evolved psychology but removes both kinship and repeated dyadic exchange from the requirements. Partner-choice or “biological market” models, developed for humans by researchers including Pat Barclay, treat cooperative reputation as a good that individuals compete to display because it makes them more attractive as future partners to the wider pool of potential cooperators, not because it earns a specific return from the specific person helped ^[12]. This mechanism does not require meeting the same partner again, unlike Trivers-style reciprocity, and it does not require selection to operate between groups, unlike the cultural group selection account; it only requires that some information about past behavior circulate and that individuals retain some freedom to choose whom they deal with — conditions a modern reputation system, credit history or reference check can satisfy for genuine strangers without either kinship or repetition.

These three accounts are not competing solutions to the same narrow question so much as different claims about where the causal weight actually sits — in an evolved, group-selected disposition to punish; in cultural and institutional invention that makes punishment cheap and coordinated; or in a market-like competition for cooperative reputation — and they are supported by different, only partially overlapping kinds of evidence. That is the honest state of the synthesis question: live, multi-sided, and not settled by any single result covered in this article.

What is settled is the fact of hypercooperation, not the mechanism behind it

Sorting what the evidence above actually supports from what it does not yields a shorter and more defensible list than the opening anomaly might suggest. Settled: humans cooperate with genetically unrelated strangers, and punish defectors at personal cost, at a scale documented across every society type tested, using instruments that reproduce the core departure from narrow self-interest in every single case studied ^{[5, 4]}. Settled: neither kin selection nor simple dyadic reciprocity can, on their own mathematics, carry cooperation to the scale of a market or a state — both mechanisms visibly run out well before they get there ^{[1, 2, 3]}. Settled: the best-documented cases of large-scale cooperation surviving for centuries involve institutions that coordinate and cheapen sanctioning, not populations of individuals independently absorbing its cost ^{[8, 11]}.

Not settled: whether laboratory costly punishment measures a disposition that actually operates to sustain cooperation outside the lab, or mainly measures something the lab’s own anonymity and stakes structure brings into being ^[8]. Not settled: the relative weight of evolved group-selected psychology, cultural-institutional invention, and partner-choice competition in explaining why that disposition, whatever produced it, holds up at all. Not settled, even, is the shape of the baseline itself — the WEIRD sampling problem means the field is still in the process of re-measuring human cooperation outside the narrow, unrepresentative slice of humanity that produced most of what was, until recently, treated as a species-wide fact ^[7].

This distinction is not academic hair-splitting; it changes what a policy or design choice can safely assume. If the operative mechanism is mainly an evolved individual disposition to punish strangers at personal cost, decentralized enforcement — peer moderation, informal sanctioning, reputation scores — should hold up reasonably well without much scaffolding around it. If it is mainly institutional invention that makes coordinated sanctioning cheap, then removing that scaffolding will not reveal a residual layer of individually costly punishment underneath; the ethnographic scarcity of uncoordinated costly punishment documented above suggests there may be nothing there to fall back on. The honest answer, on the evidence assembled here, is that both mechanisms are real, neither is fully general, and anyone designing an institution on the strength of either one alone is building on less certainty than fifty years of well-replicated experiments might seem to promise.