2. The evolution of punishment

In humans, punishment has often been studied in the context of stylised laboratory games (Chaudhuri Reference Chaudhuri2011; Guala Reference Guala2012), where individuals can pay a small monetary cost to inflict a (usually) larger fine on a target. These laboratory games have emphasised people's willingness to punish others for transgressions (Fehr and Gächter Reference Fehr and Gächter2000, Reference Fehr and Gächter2002a; Yamagishi Reference Yamagishi1986), including when they were not the victim (‘third-party punishment’; Fehr and Fischbacher Reference Fehr and Fischbacher2004). In the case of second-party punishment, where the punisher was the victim (or one victim) of the cheating individual, it is often assumed that the punisher can reap a return on investment because of a change in the target's behaviour (e.g. Clutton-Brock and Parker Reference Clutton-Brock and Parker1995), or because a bystander observes the punishment and this deters the bystander from defecting when they interact with the punisher (dos Santos et al. Reference dos Santos, Rankin and Wedekind2013). Third-party punishment is thought to be more puzzling from an evolutionary perspective, but this is largely due to the assumption that the third party will not interact with the target of punishment again in the future. This assumption – like second-party punishment in one-round games – makes it more difficult to reconcile the costly act of punishment with the possibility for downstream benefits for the punisher. Recent theoretical work has indicated that third-party punishers may nevertheless benefit when their punitive reputation is known to future interaction partners, which is likely when populations are relatively tightly structured and viscous (Roos et al. Reference Roos, Gelfand, Nau and Carr2014). Under these circumstances, the possibility for future interactions with the target of punishment or with bystanders who observe the punishment act seems to be a more important factor in understanding how punishers can be compensated for their investments, than whether the punisher was the original victim of the cheating individual or not.

Before discussing the empirical literature, it is worth bearing in mind that many of the studies we cite were conducted in the laboratory using WEIRD participants (but see for notable exceptions Ensminger and Henrich Reference Ensminger and Henrich2014; Gächter and Herrmann Reference Gächter and Herrmann2011; Henrich Reference Henrich2000; Henrich et al. Reference Henrich, Ensminger, McElreath, Barr, Barrett, Bolyanatz, Cardenas, Gurven, Gwako, Henrich, Lesorogol, Marlowe, Tracer and Ziker2010c, Reference Henrich, Boyd, Bowles, Camerer, Fehr, Gintis, McElreath, Alvard, Barr, Ensminger, Henrich, Hill, Gil-White, Gurven, Marlowe, Patton and Tracer2005, Reference Henrich, McElreath, Barr, Ensminger, Barrett, Bolyanatz, Cardenas, Gurven, Gwako, Henrich, Lesorogol, Marlowe, Tracer and Ziker2006; Herrmann et al. Reference Herrmann, Thöni and Gächter2008; Marlowe Reference Marlowe2009; Marlowe et al. Reference Marlowe, Berbesque, Barr, Barrett, Bolyanatz Cardenas, Ensminger, Gurven, Gwako, Henrich, Henrich, Lesorogol, McElreath and Tracer2008; Marlowe, et al. Reference Marlowe, Berbesque, Barrett, Bolyanatz, Gurven and Tracer2011; Wiessner Reference Wiessner2009). In contrast to the laboratory studies, data from the real world suggest that people are often unwilling to punish when opportunities to do so arise (Balafoutas et al. Reference Balafoutas, Nikiforakis and Rockenbach2014a; Baumard Reference Baumard2010; Guala Reference Guala2012; Pedersen et al. Reference Pedersen, McAuliffe and McCullough2018; but see Mathew and Boyd Reference Mathew and Boyd2011). The relatively high levels of punishment observed in laboratory games might therefore be an artefact that stems from the paucity of alternative options offered to participants (e.g. see Raihani and Bshary Reference Raihani and Bshary2015b). In real-world settings, where alternative options are available (and where people might be involved in more than one kind of social interaction with the same partner), people can withhold opportunities to help rather than actively harm transgressors (Balafoutas et al. Reference Balafoutas, Nikiforakis and Rockenbach2014a), can ‘vote with their feet’ by either expelling or distancing themselves from offenders or free-riders (Boehm Reference Boehm2001; Lee Reference Lee1979) or can use restorative justice that focuses on compensating victims for harm incurred rather than imposing penalties on the wrongdoers themselves (Hirsch et al. Reference Hirsch, Roberts, Bottoms, Roach and Schiff2003; Pupu and Wiessner Reference Pupu and Wiessner2018; Wiessner and Pupu Reference Wiessner and Pupu2012). In addition, work that has been conducted in non-industrial, small-scale societies typically finds that people living in such societies are less willing to invest in punishment of others even in the context of artificial laboratory games (Henrich et al. Reference Henrich, Heine and Norenzayan2010a; Wiessner Reference Wiessner2009) and that investment in third-party punishment in particular is less common (Marlowe et al. Reference Marlowe, Berbesque, Barrett, Bolyanatz, Gurven and Tracer2011). These caveats should be borne in mind when assessing the evidence below.

Although theoretical work has emphasised that punishment can stabilise any behaviour within a group (Boyd and Richerson Reference Boyd and Richerson1992), researchers interested in the evolution of cooperation have been especially interested in the idea that punishment can act as a tool to promote cooperation (the pathway denoted by the red arrows in Fig. 1). Under the assumption that punishment involves paying a cost to harm another individual, the higher fitness payoffs that result from the target's increased cooperation offers one plausible mechanism by which such investments can be favoured by selection (Boyd and Richerson Reference Boyd and Richerson1992; Clutton-Brock and Parker Reference Clutton-Brock and Parker1995; Raihani et al. Reference Raihani, Grutter and Bshary2010). Punishers might additionally benefit because their actions are witnessed by other individuals who then behave more cooperatively with the punisher in future (dos Santos et al. Reference dos Santos, Rankin and Wedekind2011; dos Santos and Wedekind Reference dos Santos and Wedekind2015; Jordan and Rand Reference Jordan and Rand2017; Raihani and Bshary Reference Raihani and Bshary2015b, c). Nevertheless, in laboratory games, humans willingly invest in punishment in scenarios where the act is not observed by others, where the punisher will not interact with the target again in future, and/or where the benefits will be shared by other individuals, including non-punishers. In such cases, punitive preferences could reflect a mismatch between the laboratory environment and a human psychology that evolved in environments where interactions are truly one-shot and/or unobserved (e.g. Cosmides and Tooby Reference Cosmides and Tooby1997; Tooby et al. Reference Tooby, Cosmides and Price2006, but see Fehr and Henrich Reference Fehr and Henrich2003; Hill et al. Reference Hill, Wood, Baggio, Hurtado and Boyd2014; Raihani and Bshary Reference Raihani and Bshary2015a) or could instead reflect a psychology that has been under positive selection because of the effect punishment has on group success in the context of between-group competition (e.g. Boyd et al. Reference Boyd, Gintis, Bowles and Richerson2003; blue arrows in Fig. 1).

Figure 1. Possible feedback loops by which punishment could yield inclusive fitness benefits to punishers. Red arrows denote benefits that arise because punishment causes the target or bystanders to cooperate more (with dashed lines indicating that the cooperation is directed at punisher's kin). Orange arrows denote benefits that arise because punishment changes the relative payoff difference between punisher and target. This feedback loop assumes that competition is relatively local. Blue arrows denote benefits that arise because competition between groups is stronger than competition within them. Blue arrows therefore assume that competition is relatively global.

Here, we want to highlight what we believe is a neglected strategic goal of punishment: to compete with others, rather than to encourage future cooperation. This possibility was mentioned as a footnote in Fehr and Gächter (Reference Fehr and Gächter2000) (see also Baumard Reference Baumard2010; Boehm Reference Boehm2001; Cosmides et al. Reference Cosmides, Price and Tooby2002; Nakao and Machery Reference Nakao and Machery2012; Saaksvuori et al. Reference Saaksvuori, Mappes and Puurtinen2011). It also has similarities to the theory of inequity aversion proposed by Fehr and Schmidt (Reference Fehr and Schmidt1999) as well as the levelling-down strategies used in traditional societies, discussed in Boehm (Reference Boehm2001) in that, like the competitive punishment hypothesis, both of these theories emphasise that people should be sensitive to their own payoffs (or status) relative to others’. Previous theoretical models have shown that punishment strategies that respond to relative payoffs (so-called ‘egalitarian’ punishment strategies) can be equally or even more effective than punishment strategies that rely on observing or correctly inferring the target's strategy (Scheuring Reference Scheuring2010; Tamura et al. Reference Tamura, Morita and Ihara2011) and we agree with Scheuring (Reference Scheuring2010) that the assumption that punishers do always know the strategies of players they interact with is problematic when one considers that this may not always (or even often) be the case in the real world. Additionally, a recent theoretical model on the evolution of cooperation (which was modelled as respect for resource possession) found that cooperation was stabilised when resource holders (or other individuals) responded to ‘raiding’ (attempts to steal the resource) with retaliation (i.e. punishment). Nevertheless, when the level of inequality between resource-holders and resource-seekers exceeds a critical threshold (which can be thought of as akin to variation in payoff differences among players in a laboratory game), then cooperation collapses and raiding strategies come to dominate the population until this inequality is reduced. This model, although not explicitly framed in terms of cooperation and punishment, illustrates the evolutionary logic of considering relative payoff differences in understanding the evolution of cooperation and punishment in nature.

A payoff-levelling goal might also reflect an alternative evolutionary pathway for punishment: under local competition, it could be favoured simply because it allows punishers to reduce or remove fitness differentials between themselves and targets (Cosmides et al. Reference Cosmides, Price and Tooby2002), to elevate their fitness above targets or to gain indirect fitness benefits if their kin are the main beneficiaries of the target's reduced competitiveness. The latter scenario, where the actor incurs lifetime fitness costs to impose fitness costs on another individual, fits the Hamiltonian definition of spite (Gardner and West Reference Gardner and West2004). In these scenarios (orange arrows in Fig. 1), no wrong-doing or altering of future behaviour of the target is necessary to explain the act of punishment and any behaviour change effect of punishment would be an exaptation. Importantly, the payoff consequences of acting upon punitive sentiments within the context of experimental games do not capture the direct and/or indirect fitness consequences of punishment: preferences for punishment could have been favoured by selection not because punishment tends to change targets’ behaviour but because it tends to leave the punisher (and/or his relatives) at a relative advantage. Under this view, punishment need not be restricted to those who were the victims of the original cheating individual, and could be favoured even when the punisher was an uninvolved third party.

In laboratory settings, this possibility for competition is often built into the design of the game since the cost of punishment to the punisher is usually smaller than the negative impact on the target. This confound is especially acute in studies that use variable fee-to-fine ratios (e.g. as in Denant-Boemont et al. Reference Denant-Boemont, Masclet and Noussair2007; Faillo et al. Reference Faillo, Grieco and Zarri2013; Fehr and Gächter Reference Fehr and Gächter2000; Masclet et al. Reference Masclet, Noussair, Tucker and Villeval2003; Nikiforakis Reference Nikiforakis2008a), where the punishment reduces the target's earnings by a percentage rather than a fixed sum. Under variable fee-to-fine ratios, punishers can have a larger impact on the payoffs of defectors than on the payoffs of other cooperators (as pointed out by Casari Reference Casari2005).

Below we outline the key predictions made by the hypothesis that punishment is (primarily) a tool to promote cooperation, along with a brief summary of the theoretical and empirical evidence that supports or undermines each prediction.

3. Brief summary of findings

The findings above indicate that: (i) punishers are not always cooperative; (ii) being punished does not consistently have the expected effect of increasing cooperation from the target, or of increasing group payoffs; (iii) cooperators as well as defectors can be targeted for punishment; and (iv) punishers are often proximately motivated by payoff differences rather than by cheating or losses per se. These findings are problematic from a perspective that views punishment solely as a tool to promote cooperation but are consistent with the hypothesis that punishment serves a competitive function. The fact that punishment can be associated with different strategic goals also helps to illuminate how punishers are perceived. The aims of general deterrence and a desire to create equal outcomes among all players (egalitarianism) are both, broadly speaking, aims that are consistent with producing public goods. Thus, if people infer that punishment has these collective goals in mind, then they should generally approve of punishers. In contrast to this prediction, most published studies indicate that punishers are generally disliked and disapproved of (reviewed by Raihani and Bshary Reference Raihani and Bshary2015c), and that punishers might only be approved of in a very restricted range of circumstances where competitive or self-serving motives can be effectively ruled out (see Barclay Reference Barclay2006; Jordan et al. Reference Jordan, Hoffman, Bloom and Rand2016; Kiyonari and Barclay Reference Kiyonari and Barclay2008; Raihani and Bshary Reference Raihani and Bshary2015b, c). We suggest that punishers may often be disliked or disapproved of because punishment is also often consistent with a competitive strategic aim, whereby the punisher seeks to increase their own payoffs relative to those of the target.

The possibility that punishment can result in immediate fitness benefits to punishers – without the need for any conditional response from targets – somewhat blurs the distinction between punishment and what we (and others) had previously labelled ‘sanctions’ (interactions where aggressors derive immediate benefits from their actions, also known as negative pseudo-reciprocity; Bergmuller et al. Reference Bergmuller, Johnstone, Russell and Bshary2007; Raihani et al. Reference Raihani, Thornton and Bshary2012a) or negative indirect reciprocity Bhui et al. Reference Bhui, Chudek and Henrich2019). Under negative pseudo-reciprocity (sanctions), investments in harming another individual can be self-serving even if the target does not change his behaviour. For example, female coral gobies have a linear size-based dominance hierarchy and may evict any subordinate who grows too close in size to them (Wong et al. Reference Wong, Buston, Munday and Jones2007). Previously we argued that this did not fit the definition of punishment, as the evictor gains a direct benefit when she removes a competitor that does not rely on the competitor behaving more cooperatively in future interactions with the evictor (Raihani et al. Reference Raihani, Thornton and Bshary2012a). However, it is now apparent that many examples of punishing in humans also have the same property: despite involving a short-term payoff-reduction, punitive strategies can be fitness-enhancing if they change the relative payoffs of punishers and their targets. Many ethnographic accounts describe how deviants or cheats risk being ostracised or deserted by fellow group-members, and there are also many examples of homicidal or otherwise dangerous individuals being executed by other members of the group (Boehm Reference Boehm2011; Wrangham Reference Wrangham2019). These kinds of harmful acts that also terminate the interaction between cheats and victims would involve similar payoff structures as eviction in the goby example above. To account for the variety of potential direct and indirect fitness benefits, we propose that the definition of punishment should focus on the act itself: an act that lowers the immediate payoff of actor and recipient, rather than the circumstances that trigger punishment or the consequences in terms of the target's behaviour.

4. Punishment may be rare, but punitive preferences are not

A competitive account of punishment can potentially help to reconcile the fact that punishment of the variety we study in laboratory settings seems to be relatively rare in the real world, while accounting for the expression of punitive sentiment that is more frequently observed. For instance, ethnographies suggest that a common expression of punitive sentiment occurs in the form of mild criticism or mocking the target (Baumard Reference Baumard2010; Boehm Reference Boehm2011; Wiessner Reference Wiessner2005). For example, Hoebel (Reference Hoebel1954, p. 93) described how conflicts among Inuit could be resolved with song-duels, where accused and accuser took it in turns to ridicule one another in verse. Similarly, Ju/’Hoansi bushmen mediate conflict and express displeasure with others by using various forms of reputational put-downs, including mocking, ridicule and outright criticism, rather than physical aggression (Wiessner Reference Wiessner2005). A parallel example comes from studies of the Cheyenne, where shaming of thieves occurs by a statement to the effect of ‘If I had known you wanted that thing, I would have given it to you’ (Hoebel Reference Hoebel1954, p. 169). Notably, it is not only defectors or cheats who are targeted for shaming and ridicule but also individuals who have or seek status or prestige. Indeed, Wiessner (Reference Wiessner2005) writes that a key function of mocking and ridicule is to ‘level big-shot behaviour’. Perhaps because cooperative acts are associated with reputational gains for the cooperative individual, ostensive or showy acts of generosity are often hidden (Raihani Reference Raihani2014), ignored (Wiessner Reference Wiessner2005) or even ridiculed (e.g. ‘do-gooder derogation’; Minson and Monin Reference Minson and Monin2012), findings which also speak to the competitive underbelly of punitive sentiment.

Other ways that punishers can inflict reputation or status costs on targets is via gossip (Beersma and Van Kleef Reference Beersma and Van Kleef2012; Dunbar Reference Dunbar2004). Gossip is prevalent in real-world settings and seems to have similar payoff structure to punishment in that it is costly to be perceived as a gossip (Adams and Mullen Reference Adams and Mullen2012) and to be gossiped about (Feinberg et al. Reference Feinberg, Willer, Stellar and Keltner2012). Gossip typically refers to conversation or statements that convey social information about absent third parties. A disproportionate amount of human speaking time (~65%, with little variation across cultures, ages or gender; Dunbar Reference Dunbar2004) is used to gossip. Gossip can be used to convey or gather information about interaction partners in the absence of direct observation and individuals are sensitive to the potential effects of gossip on their reputation, behaving more cooperatively when reputation is at stake (Feinberg et al. Reference Feinberg, Willer, Stellar and Keltner2012, Reference Feinberg, Willer and Schultz2014; Jolly and Chang Reference Jolly and Chang2018; Wu et al. Reference Wu, Balliet and Van Lange2016a, Reference Wu, Balliet and Van Langeb).

While gossip can in theory transmit either positive or negative information about a target, humans apparently prefer to gossip negatively (Feinberg et al. Reference Feinberg, Willer and Schultz2014; Peters et al. Reference Peters, Jetten, Radova and Austin2017). Like punishment, negative gossip is subjectively rewarding to gossipers and is elicited by the same negative emotions (i.e. frustration and anger at interacting with a cheat) that produce punishment (Feinberg et al. Reference Feinberg, Willer, Stellar and Keltner2012). As with punishment, it seems that there is also the potential for gossip to stem from different strategic aims. For instance, Feinberg et al. (Reference Feinberg, Willer, Stellar and Keltner2012) show that tendency to gossip is associated with prosocial value orientation and people state that the reason for gossip is to help warn the recipient about the previous untrustworthy behaviour of a confederate. However, others have shown that the tendency to gossip negatively is predicted by personality traits implicated in antisocial preferences, including psychopathy and Machiavellianism (Lyons and Hughes Reference Lyons and Hughes2015). In her study of the Ju/’hoansi bushmen, Wiessner (Reference Wiessner2005, Reference Wiessner2009) found that ~30% of cases of gossip were ‘unfounded and concocted out of jealousy or social strategies’ (P. Wiessner, personal communication). Similarly, empirical work has shown that gossip is most likely to occur in triadic settings, with two people gossiping about a higher-status other (Ellwardt et al. Reference Ellwardt, Wittek and Wielers2012). As with punishment, it is unlikely that gossip functions to induce cooperation from the target towards the gossiper in the future (i.e. special deterrence), although it might be the case that gossipers are motivated by a more general deterrent aim. One might additionally posit that gossipers, like punishers, are motivated by retributive or competitive aims (i.e. to damage a defector's reputation to increase gossiper's relative status and improve the gossiper's relationships with the audience, e.g. Bosson et al. Reference Bosson, Johnson, Niederhoffer and Swann2006; Jolly and Chang Reference Jolly and Chang2018; Peters et al. Reference Peters, Jetten, Radova and Austin2017). These hypotheses are amenable to empirical testing.

6. Institutionalising punishment

Given the difficulty in correctly pinpointing why someone punishes, and what their strategic aim might be, it is perhaps unsurprising that we do not typically see participants in experimental games responding to punishment by increasing cooperation. We suggest that the way people respond to punishment is likely to hinge fundamentally on the motives and aims they attribute to punishers – and that these attributions will hinge crucially on the features of the interaction – broadly speaking, whether the punisher can potentially derive personal benefits from their actions. As far as we know, the link between features and attributions has seldom been explored in empirical studies and thus represents a fruitful avenue for inquiry (see Ho et al. Reference Ho, Cushman, Littman and Austerweil2019 for a first step in this direction). The tendency to infer that punishers are driven by cooperative or competitive motives is expected to vary with the context of punishment (specifically whether the context allows that the punisher can benefit from their actions, for example, if punishers expect to interact with targets in future, or can improve their own relative payoffs through the act of punishing – this is discussed at length in Raihani and Bshary Reference Raihani and Bshary2015c). The tendency to attribute competitive motives to punishers might also vary across individuals. For instance, previous research has shown that there is enormous variation in the general population in the tendency to attribute malevolent intentions to others in ambiguous social settings (Raihani and Bell Reference Raihani and Bell2017a; Saalfeld et al. Reference Saalfeld, Ramadan, Bell and Raihani2018) and that the tendency to attribute malevolent intentions to others also predicts punitive responses in social interactions (Raihani and Bell Reference Raihani and Bell2017b). In the standard laboratory setting, where individuals receive punishment in the form of monetary fines without any explanation for the basis of the punishment nor agreement upon the actions that warrant punishment and how severe punishment should be, there is enormous scope for variation and error when targets attempt to interpret a punisher's intentions. For instance, targets might infer that the punisher dislikes interacting with a cheat and wants them to cooperate in the next round. However, a target of punishment might also reasonably infer that the punisher simply wishes to inflict harm rather than to promote cooperation. The intentions attributed to punishers might also vary with societal norms surrounding the use of punishment. For instance, in societies where individuals expect others to cooperate, punishment might be viewed as a legitimate response to defection, with the inferred intention being that punishers seek to change defectors’ behaviour. In societies where such norms do not exist, punishers might be more likely to be perceived as having competitive aims, with the result that punishment is less likely to foster cooperation (see discussion in Balliet and Van Lange Reference Balliet and Van Lange2013). Empirical work addressing whether intention attribution affects responses to punishment in economic games would be useful to explore if and how intention attribution affects targets’ responses to punishment.

Suggestive evidence that the perceived intentions of punishers do affect targets’ responses comes from empirical studies where punishment can be construed as being driven to a greater or lesser extent by self-interested motives: in a two-player Trust Game, trustees who demanded a large investment from the partner under the threat of punishment received less than those who did not issue a punishment threat (Fehr and Rockenbach Reference Fehr and Rockenbach2003; see also Houser et al. Reference Houser, Xiao, McCabe and Smith2008). Other work indicates that people attempt to infer the proximate motives underpinning punishment decisions and respond accordingly. Xiao (Reference Xiao2013) used a sender–receiver game, where the sender could benefit by deceiving the receiver, and introduced a punishment stage where the receiver could additionally punish the sender. In some treatments, receivers immediately benefitted from punishing the partner, whereas in others punishment was associated with a payoff reduction. When receivers could benefit from punishing the sender, third-party bystanders were less likely to infer that the sender had cheated, compared with cases where receivers could not benefit from punishing the sender (Xiao Reference Xiao2013). Similarly, a meta-analysis finds that decisions to reduce someone else's payoffs are more likely to promote cooperation when these payoff reductions are costly to the actor, rather than free, to implement (Balliet et al. Reference Balliet, Mulder and Van Lange2011).

Generally, we predict that the tendency to cooperate in response to punishment will be inversely related to the extent to which targets of punishment can infer that the punisher has competitive aims. Cooperating or acquiescing in response to competitive punishment can be construed as capitulating in response to aggression from an evenly matched peer (as punishment is typically operationalised in experimental economic games). Instead, under natural conditions, there might be selection against targets allowing themselves to be subordinated, and individuals might attempt to preserve their status by retaliating against evenly matched aggressors (see Boehm Reference Boehm2011). Indeed, the rejection of unfair Ultimatum Game offers has been interpreted as a means of avoiding acquiring a subordinate status (Yamagishi et al. Reference Yamagishi, Horita, Mifune, Hashimoto, Li, Shinada, Miura, Inukai, Takagishi and Simunovic2012) and punishment is also linked to testosterone levels in men (Burnham Reference Burnham2007), as would be expected if punishment is viewed as a dominance competition rather than as a way to communicate cooperative norms. Selection against subordination could be stronger if capitulating to aggressors is also observed by others. In much the same way as gaining a punitive reputation can yield an advantage to punishers (Barclay Reference Barclay2006; dos Santos et al. Reference dos Santos, Rankin and Wedekind2011, Reference dos Santos, Rankin and Wedekind2013; dos Santos and Wedekind Reference dos Santos and Wedekind2015; Raihani and Bshary Reference Raihani and Bshary2015b, c), a reputation for being easily subordinated might be disadvantageous to acquire (see Cohen et al. Reference Cohen, Nisbett, Bowdle and Schwarz1996; Crombag et al. Reference Crombag, Rassin and Horselenberg2003; Osgood Reference Osgood2017). This hypothesis deserves further theoretical and empirical attention.

The possibility for punishment to spark retaliation, rather than cooperation, is prevalent in historical (Barrett et al. Reference Barrett, Harrison and Harrison2005) and ethnographic records (Boehm Reference Boehm2011; Hoebel Reference Hoebel1954; McCullough et al. Reference McCullough, Kurzban and Tabak2013; Wiessner Reference Wiessner2005; see Jackson et al. Reference Jackson, Choi and Gelfand2019 for a recent review). Indeed, many societies seem to have independently converged on customs (e.g. duels and other ritualised contests) that circumscribe the contexts in which punishment can be administered, perhaps to limit these detrimental consequences (Boehm Reference Boehm2011). As a consequence, decentralised peer punishment of the variety studied in experimental economic games is often rare or absent among humans in the real world (Guala Reference Guala2012; Pedersen et al. Reference Pedersen, McAuliffe and McCullough2018; Wiessner Reference Wiessner2005), although it tends to be more common in contexts where such customs are lacking (reviewed in Jackson et al. Reference Jackson, Choi and Gelfand2019). These findings speak to a more specific insight, which is that for punishment to be effective as a tool to convert cheaters into co-operators, it must be perceived as legitimate (Baldassarri and Grossman Reference Baldassarri and Grossman2011; Bowles and Gintis Reference Bowles and Gintis2013; Ertan et al. Reference Ertan, Page and Putterman2009; Faillo et al. Reference Faillo, Grieco and Zarri2013; Gross et al. Reference Gross, Méder, Okamoto-Barth and Riedl2016; Tyler Reference Tyler2006; Villatoro et al. Reference Villatoro, Andrighetto, Brandts, Nardin, Sabater-Mir and Conte2014; Xiao and Tan Reference Xiao and Tan2014; Zheng and Nie Reference Zheng and Nie2013). Here, we define legitimisation as any process, mechanism or institution that reduces the scope for punishment to be driven by competitive aims and thereby increases the probability that targets cooperate in response to punishment.

Perhaps the simplest legitimising mechanism is communication: allowing people to explain to targets why they are punishing means that cooperation is more likely to ensue (Janssen et al. Reference Janssen, Holahan, Lee and Ostrom2010). Legitimacy can also be promoted by more formalised agreements or customs, which are often collectively referred to as institutions (Cushman Reference Cushman2015; Hurwicz Reference Hurwicz1996; North Reference North1990; Powers et al. Reference Powers, van Schaik and Lehmann2016). Specifically, institutions can be thought of as political game forms which change the payoffs associated with social interactions (Hurwicz Reference Hurwicz1996; Powers et al. Reference Powers, van Schaik and Lehmann2016). Institutions probably play a vital role in changing the game form of punishment from one where individuals can benefit by harming others, regardless of how the target behaves (i.e. as is the case with competitive punishment) to one where it only pays to punish cooperative norm violators (Boyd Reference Boyd2017). In much the same way, it has been posited that institutions played (and continue to play) a vital role in transforming social dilemmas (where the strategic incentive is to defect) into games where the individually self-interested strategy is to cooperate (Boyd Reference Boyd2017; Powers et al. Reference Powers, van Schaik and Lehmann2016; Sigmund et al. Reference Sigmund, De Silva, Traulsen and Hauert2010). Although institutions might thus generate benefits for all members of a group, we also note that the historical record contains many examples of institutions being used to further the interests of a powerful few at the expense of the many (Acemoglu Reference Acemoglu2006; Acemoglu et al. Reference Acemoglu, Johnson and Robinson2001; Briggs et al. Reference Briggs, Briggs, Harrison, McInnes, Vincent, Briggs, Harrison, McInnes and Vincent2005).

Crucially, institutions have often been omitted from standard experimental economic investigations of peer punishment, which might help to explain the relatively mixed results on the efficacy of peer punishment in these stylised settings. One way to legitimise punitive acts is for participants to self-select into groups with punishment regimes (see Gürerk et al. Reference Gürerk, Irlenbusch and Rockenbach2006 for experimental evidence and Frey and Sumner Reference Frey and Sumner2019 for a real-world example), to vote for sanctioning institutions (Sutter et al. Reference Sutter, Haigner and Kocher2010) or to be involved in shaping institutions that dictate how punishment will be implemented (Decker et al. Reference Decker, Stiehler and Strobel2003; Ertan et al. Reference Ertan, Page and Putterman2009). Recent experimental work indicates that the requirement for consensus is an important institution for legitimising punishment: when punishment decisions are reached by consensus, punishment is more effective at promoting cooperation (Cardenas Reference Cardenas2000; Casari and Luini Reference Casari and Luini2009; Eriksson et al. Reference Eriksson, Strimling, Andersson, Aveyard, Brauer, Gritskov, Kiyonari, Kuhlman, Maitner, Manesi, Molho, Peperkoorn, Rizwan, Stivers, Tian, Van Lange, Vartanova, Wu and Yamagishi2017; Ertan et al. Reference Ertan, Page and Putterman2009; Hilbe et al. Reference Hilbe, Traulsen, Röhl and Milinski2014; Pfattheicher et al. Reference Pfattheicher, Bohm and Kesberg2018; Shinada and Yamagishi Reference Shinada and Yamagishi2007; Villatoro et al. Reference Villatoro, Andrighetto, Brandts, Nardin, Sabater-Mir and Conte2014; Zheng and Nie Reference Zheng and Nie2013). Many laboratory studies reporting positive effects of punishment on targets’ subsequent behaviour (e.g. Fehr and Gächter Reference Fehr and Gächter2000, Reference Fehr and Gächter2002a; Pfattheicher et al. Reference Pfattheicher, Bohm and Kesberg2018; Rockenbach and Milinski Reference Rockenbach and Milinski2006; Shinada and Yamagishi Reference Shinada and Yamagishi2007) do not report how many people independently punished a cheating player, meaning that it is unclear to what extent any positive effects of punishment on the target's subsequent cooperation were driven by these seemingly consensus-based cases in which two or three peers decided independently to punish the same target.

Data from real-world settings also support the idea that consensus-based legitimacy can increase the efficacy of punishment. For example, Mathew and Boyd (Reference Mathew and Boyd2011) argue that punishment is only administered after a consensus-making process, noting that ‘vigilante’ punishment that is administered without prior consensus is disapproved of. Similar historical examples also exist. In medieval England, men were organised into small groups called ‘tithings’ who were responsible for administering punishment to any defector in their group. Interestingly, these real-world examples show that participation in group punishment is often (i) forced rather than voluntary and (ii) administered by the allies or kin of the target (Hoebel Reference Hoebel1954, p. 89). Both features might mean that there is less scope for targets to infer that punishers are motivated by competitive aims and/or benefit via enhancing their own relative payoffs in such interactions (see also Boehm Reference Boehm2011 for ethnographic examples). Moreover, consensual punishment administered by group might reduce the scope for retaliation (and feuds) simply by virtue of the cost associated with retaliating against a larger group. Legitimacy is also helped by restricting who can punish: when people can only punish those who contribute less than them (Faillo et al. Reference Faillo, Grieco and Zarri2013; Grieco et al. Reference Grieco, Faillo and Zarri2017), it reduces the possibility that punishment is perceived as a competitive action.

Outsourcing punishment to authorities (also called ‘centralising’ punishment) might be another way to increase legitimacy. When given the choice, people prefer to pay taxes to provide centralised ‘pool’ punishment rather than peer punishment (Andreoni and Gee Reference Andreoni and Gee2012; Traulsen et al. Reference Traulsen, Röhl and Milinski2012), although other work finds that, when given the choice between a punishment or no punishment regime, many people at least initially opt for the latter (Guillen et al. Reference Guillen, Schwieren and Staffiero2007; Gürerk et al. Reference Gürerk, Irlenbusch and Rockenbach2006). Moreover, when a centralised punishment mechanism is available, people reduce their investments in peer punishment, leading to overall higher average payoffs (mainly because people are not paying for peer punishment and retaliation is reduced in the centralised punishment regime; Andreoni and Gee Reference Andreoni and Gee2012). Nevertheless, evidence that centralised punishment is more effective than peer (decentralised) punishment is mixed (Andreoni and Gee Reference Andreoni and Gee2012; Baldassarri and Grossman Reference Baldassarri and Grossman2011; Carpenter et al. Reference Carpenter, Kariv and Schotter2012; Grieco et al. Reference Grieco, Faillo and Zarri2017; Nosenzo and Sefton Reference Nosenzo and Sefton2014; O'Gorman et al. Reference O'Gorman, Henrich and Van Vugt2009; Traulsen et al. Reference Traulsen, Röhl and Milinski2012), with a meta-analysis indicating that centralised punishment is actually less effective than decentralised punishment at promoting cooperation (Balliet et al. Reference Balliet, Mulder and Van Lange2011). We predict that the perceived legitimacy of the authority will be crucial for determining whether the punishment they deliver is effective, and thus increases group welfare. For example, an experiment in Uganda found that participants were more than twice as likely to increase cooperation in response to punishment from an elected monitor than in response to punishment from a monitor who was randomly assigned to the role (Baldassarri and Grossman Reference Baldassarri and Grossman2011). Often, individuals with the power to wield punishment are either elected to that role on the basis of their character and can be demoted if they are perceived to be corrupt (Lierl Reference Lierl2018) or are held in check by a variety of mechanisms (institutions) that preclude or strongly disincentivise corrupt leadership methods (Bøggild and Petersen Reference Bøggild and Petersen2016).

We also predict that centralised punishment will be more effective where there is little potential for authorities to obtain personal benefits from their punitive acts. Sometimes, authorities can benefit from the punishments that are administered – for example, Hoebel (Reference Hoebel1954) describes how Inuit shamans acting as authorities in disputes involving women would often instruct the woman to have intercourse with him (his powers being conveniently able to counteract the effects of her sinning). Similarly, in many small-scale societies, monitoring is incentivised by allowing individuals who detect crimes to keep a portion of any fine that is levied on a cheat (Ostrom Reference Ostrom2015). In the Ifugao of Luzon, people known as monkalum act as informal go-betweens in disputes and conflicts, and benefit in this quasi-judiciary role from taking fees from defendants (Hoebel Reference Hoebel1954, p. 116). In such situations, we expect that even punishment from authorities will be scarcely better than conventional peer punishment in promoting cooperation. Punishment is more likely to have a cooperation-enforcing effect when authorities are perceived to be impartial (Muthukrishna et al. Reference Muthukrishna, Francois, Pourahmadi and Henrich2017). Indeed, a recent experiment conducted in Liberia found that punishment from leaders only promoted cooperation when leaders received a flat fee for the role of punisher, rather than benefitting from administering the punishment itself (Beekman et al. Reference Beekman, Nillesen and Voors2018). Similarly, experimental evidence from real-world leaders of a forest commons management programme in Ethiopia revealed that: (i) leaders varied in their propensity to punish indiscriminately (i.e. competitively) in an experimental setting; and (ii) competitive leaders were less effective at promoting collective cooperation in the context of the forest management (Kosfeld and Rustagi Reference Kosfeld and Rustagi2015). Procedural fairness thus seems to be an important device for increasing legitimacy (Bøggild and Petersen Reference Bøggild and Petersen2016; Cremer and van Knippenberg Reference Cremer and van Knippenberg2003; Eisner et al. Reference Eisner, Murray, Ribeaud, Averdijk and van Gelder2017).

7. Moving forwards

We believe that new theoretical and empirical studies are now needed to examine the conditions under which punishment with different strategic aims outlined above (Table 1) could be favoured by selection. In this endeavour, we suggest that identifying the relevant ecological and social factors that affect the use and efficacy of punishment will be crucial. These factors are likely to include (but are not limited to) population structure, resource availability, the likelihood of repeated interactions, the scale of competition and the strength and content of the institutions surrounding the acceptable use of punishment. We predict that competitive punishment should be increasingly favoured by selection as the scale of competition becomes increasingly local, although we also believe that more work is currently needed on how ecology (resource availability, population structure, pathogen prevalence, environmental predictability, etc.) affects the scale of competition and the form and content of institutions across societies. Ideally, theoretical models would also start to explore the evolutionarily stable responses to punishers (as in Hilbe and Traulsen Reference Hilbe and Traulsen2012; Morris et al. Reference Morris, MacGlashan, Littman and Cushman2017). For example, it would be helpful to know under what conditions cooperating in response to punishment might be under negative selection (readily submitting to peers might be under negative selection when observed by others, for example).

Table 1. Strategic aims underpinning punishment and the problematic findings from the published literature that are inconsistent with each goal

In empirical studies, we believe it is important that punishment is implemented using various cost ratios (e.g. as in Bone and Raihani Reference Bone and Raihani2015; Bone et al. Reference Bone, Wallace, Bshary and Raihani2015, Reference Bone, Wallace, Bshary and Raihani2016; Egas and Riedl Reference Egas and Riedl2008; Nikiforakis and Normann Reference Nikiforakis and Normann2008; Sefton et al. Reference Sefton, Shupp and Walker2007). We are not aware of any biological rationale for the commonly implemented 1:3 fee-to-fine ratio. In fact, giving each person in a dilemma this same technology makes little biological sense. A 1:3 ratio implies a dominance hierarchy or a power asymmetry (it costs me a small amount to inflict a larger cost on you): it is not obvious how an individual can be both a dominant and a subordinate in such a power struggle. More generally, we also note that the standard 1:3 fee-to-fine ratio has the competitive function inbuilt by default. While some authors have argued that low-cost, high-impact punishment is necessary to promote cooperation (Chaudhuri Reference Chaudhuri2011; Nikiforakis and Normann Reference Nikiforakis and Normann2008; but see Balliet et al. Reference Balliet, Mulder and Van Lange2011), it is also possible that these competitive fee-to-fine ratios are more likely to prompt retaliation, thereby undermining cooperation. We suggest that work exploring punishment that is more costly to administer than it is to receive might be more likely to be interpreted as an honest signal of discontent (rather than as an expression of competitive motives) and might thus be more likely to induce targets to cooperate (see Balliet et al. Reference Balliet, Mulder and Van Lange2011 for evidence on the importance of punishment costs for promoting cooperation).

An underexplored line of research is to compare punishment in humans explicitly with that in other species. Few studies have adapted typical human laboratory games to test for punishment decisions in primates (Jensen et al. Reference Jensen, Call and Tomasello2007a, Reference Jensen, Call and Tomasellob; Leimgruber et al. Reference Leimgruber, Rosati and Santos2016), and even rarer are studies that explore to what extent punishment may stabilise human cooperation under conditions in which punishment stabilises cooperation in other species (but see Bone et al. Reference Bone, Wallace, Bshary and Raihani2016). One important theme in the cleaner fish mutualism seems to be asymmetries. Among cleaners and their clients, there are clear asymmetries in strategic options owing to pre-determined roles: the cleaners are service providers that can cheat while the clients are what the name implies, lacking options to cheat in return (unless they are predators and hence could eat the cleaner; Trivers Reference Trivers1971). In the case of male cleaners punishing cheating female partners, there are asymmetries in power among interacting players as male cleaners are larger than and dominant to female partners. Indeed, females never punish males for cheating a shared client. In the vervet monkey example, punishment seems to be used as a form of negotiation over what actions to take, as females have more to gain than males from escalating and attacking the neighbouring group. Thus, it appears that intergroup encounters are not about cooperating and free-riding but about diverging interests between the sexes. Capturing key features of these animal examples in new laboratory-based experiments may yield important insights about punishment in humans, no matter whether results will be similar or different from the animal cases (Bone et al. Reference Bone, Wallace, Bshary and Raihani2016).

We also need more studies which tease out possible motives underpinning punishment decisions and responses to punishment. For example, if punishment serves a competitive function, then helpful individuals who reap disproportionate payoffs from their actions should be more likely to be punished than unhelpful individuals who nevertheless remain at a relative disadvantage – and individuals might therefore avoid any actions (not just cheating) that leave them at a payoff advantage over partners in social interactions.

To conclude, we suggest that human punishment might sometimes or by some persons be aimed at changing the behaviour of targets but that this is not always the case – and that competitive motives are understudied and yet arguably more important. There is more evidence to suggest that the possibility of punishment acts as an antecedent deterrent against would-be cheats than as a tool to convert cheats into cooperators – but this could be an exaptation, with the primary function of punishment being to reduce personal payoff disadvantages or improve punisher's own payoffs relative to the target's. Certainly, the existing empirical evidence on the proximate motives underpinning punishment decisions suggests that punishment decisions frequently stem from fairness and/or status concerns, rather than from the desire to reciprocate harm. Researchers interested in the role of punishment in supporting human cooperation – including us! – need to start thinking hard about additional possible functions of punishment in humans.