Does cultural group selection make unambiguous predictions about human behaviour?

There are some core elements of CGS that make it an evolutionary process which plausibly explains elements of human cooperation (Richerson et al. Reference Richerson, Baldini, Bell, Demps, Frost, Hillis, Mathew, Newton, Naar, Newson, Ross, Smaldino, Waring and Zefferman2016): (a) cultural variation is responsible for some behavioural differences between groups; (b) certain cultural traits are inherited and persist over time; and (c) different cultural traits are responsible for the success of certain cultural groups. In some cases the evidence for CGS is irrefutable: larger and more organised groups with better technology have outcompeted and displaced smaller and less organised groups with simpler technology throughout human history (Currie and Mace Reference Currie and Mace2009; Diamond Reference Diamond1997). Examples include agriculture replacing hunting-and-gathering, the colonisation of the Americas and the rise of nation states. Despite this broad high-level agreement, in discussions and models of CGS there are different varieties of CGS which can result in this process and need to be separated. These often make different predictions regarding human behaviour and altruism.

Some presentations of CGS claim that it provides an explanation for altruism in humans (Henrich Reference Henrich2004a; Richerson and Boyd Reference Richerson and Boyd2005). These explanations focus on mechanisms of social learning, such as conformism or prestige-bias, which have evolved because they are adaptive in most contexts but may be maladaptively applied when acquiring cooperative behaviour. Taking an example from Richerson and Boyd (Reference Richerson and Boyd2005, pp. 162–163), consider two groups linked by migration: one group is predominantly religious and the other is non-religious. In the non-religious group, no-one helps anyone. In the religious group, individuals altruistically help one another. Within each group, migrants adopt the prevailing behaviour via conformism. However, individuals also have a preference to adopt the selfish non-religious behaviour, meaning that in the religious group there will always be a proportion of non-religious free-riders who have greater fitness than religious individuals. Thus, being religious is both KS- and ML-altruistic, as individuals would have greater fitness if they were non-religious. However, owing to conformism these group differences can persist, and religious groups will outcompete non-religious groups in between-group competition (for a formal model of this process, see chapter 7 of Boyd and Richerson Reference Boyd and Richerson1985). Richerson and Boyd (Reference Richerson and Boyd2005) label such behaviours ‘maladaptive adaptations’, since they rely on constraints on social learning strategies which can lead to maladaptive behaviour. As such, I will call these explanations ‘maladaptive CGS’ (while noting that, on average, these strategies may be adaptive for acquiring traits other than cooperation).

Maladaptive CGS was invoked in many early discussions of CGS. As Richerson and Boyd (Reference Richerson and Boyd2005) state, ‘if cultural rules arise [via conformism] that cause individuals to sacrifice their own interests for the good of the group, group selection can cause the frequency of individually costly but group-beneficial traits to increase’ (p. 162), and that CGS is an ‘engine for generating maladaptations from the narrow genes’-eye point of view’ (p. 244). Note that not all models of CGS based on social learning result in maladaptive behaviour; pay-off-biased imitation – where individuals preferentially copy fitness-enhancing traits – can also lead to the spread of cooperative behaviour, either by imitating the behaviour of successful groups (Boyd and Richerson Reference Boyd and Richerson2002) or by migrating to successful groups and then adopting the local behaviour (Boyd and Richerson Reference Boyd and Richerson2009b). However, when applied within groups, pay-off-biased transmission will frequently not lead to stable cooperative outcomes, as defectors have greater fitness than cooperators within groups, so pay-off-biased transmission will lead to these selfish behaviours being copied (van den Berg, Molleman, and Weissing Reference van den Berg, Molleman and Weissing2015).

Rather than focus on social learning mechanisms to explain cooperative behaviour, other, often more recent, discussions of CGS tend to focus on mechanisms of reward, reputation, punishment and a norm psychology; what I will dub ‘normative CGS’. From this perspective, behaviour is not altruistic as cooperators would not have lower fitness than defectors within groups (Boyd Reference Boyd2017; Boyd and Richerson Reference Boyd and Richerson2009a). Take punishment: if all individuals in a group punish norm-breakers (and punish those who do not punish norm-breakers), punishment can stabilise any behaviour within groups (Boyd and Richerson Reference Boyd and Richerson1992). As non-punishers would have lower fitness, this behaviour is not maladaptive or altruistic by either kin selection or MLS definitions. As Boyd (Reference Boyd2017) recently stated, CGS ‘is about selection among groups with different social arrangements that are evolutionarily stable, not about the evolution of individually costly group-beneficial behavior’ (p. 107; emphasis in original) and that ‘cultural group selection predicts that norms maintained by self-interest will tend to be group beneficial, not that people will be altruists’ (p. 189). However, other authors do claim that such punishment of norm-breakers is against self-interest, and therefore altruistic (Bowles and Gintis Reference Bowles and Gintis2011). Note that normative and maladaptive CGS can work in concert, such as a conformist bias to punish norm-breakers (Henrich and Boyd Reference Henrich and Boyd2001), but I largely treat them as separate processes for conceptual clarity and because they make qualitatively different predictions regarding human cooperative behaviour.

Rather than focus on mechanisms to stabilise group behaviour followed by equilibrium selection, other formulations of CGS emphasise the conflict between within- and between-group selection, without necessarily requiring stable group behaviour (Sober and Wilson Reference Sober and Wilson1998; Waring et al. Reference Waring, Kline, Brooks, Goff, Gowdy, Janssen, Smaldino and Jacquet2015), which I will call ‘mechanism-neutral CGS’. For instance, Turchin et al. (Reference Turchin, Currie, Turner and Gavrilets2013) and Smaldino (Reference Smaldino2014) define CGS (what they call ‘cultural multi-level selection’) as the process by which complex and otherwise-costly group-beneficial cultural institutions can proliferate over other institutions owing to between-group competition, especially warfare (by ‘costly’, I mean that the within-group term is negative, but the trait is maintained via its benefits in between-group competition; it does not imply that the trait is KS-altruistic). From this mechanism-neutral CGS perspective, social learning mechanisms and/or a norm psychology to stabilise group behaviour are not necessary as the proximate mechanisms are left unspecified. Rather, costly institutions are maintained because of inter-group competition; without group competition, these institutions would probably become simpler and less costly. Although mechanism-neutral CGS is similar to maladaptive and normative CGS in many respects, in that all require group competition for group-beneficial behaviour to spread, a crucial difference is that the latter do not require group competition for cooperation to be sustained, as cooperation is stable within groups. As mechanism-neutral CGS does not rely on mechanisms to stabilise between-group differences, group competition is required to maintain costly institutions (this also means that groups need not be in equilibrium, and therefore that mechanism-neutral CGS does not rely on equilibrium selection). Mechanism-neutral CGS may be weakly ML-altruistic, but it does not require KS-altruism, strong ML-altruism or that behaviour is maladaptive. Unless otherwise specified, when mentioning ‘CGS’ below I am referring to maladaptive and normative CGS, rather than mechanism-neutral CGS.

All varieties of CGS presented above agree that competition between cultural groups is important, but they differ in the proximate mechanisms and processes by which this is achieved and whether the resultant behaviour is altruistic or self-interested (Table 1). These conflicting definitions and differing behavioural expectations can make testing prediction of CGS difficult. Many critiques of CGS focus on maladaptive CGS and argue that content-blind mechanisms of social learning (e.g. conformism) are unlikely to explain human cooperation (e.g. Lamba and Mace Reference Lamba and Mace2011). This conclusion is likely to be true. However, as CGS does not only rely on blind social learning mechanisms, the dismissal of CGS based on a rejection of maladaptive CGS may be rather uncharitable. At the same time, a proliferation of theories which fall under the name ‘cultural group selection’, yet make different predictions regarding human cooperative behaviour, is almost certain to foster confusion.

Table 1. Summary of predictions and proposed proximate mechanisms made by different versions of cultural group selection (CGS).

^a In the religious vs non-religious group example presented in the main text, there is conflict between the levels of selection, as in the religious group the minority of non-religious individuals has greater within-group fitness than the religious individuals. Despite within-group selection acting against religious individuals, conformism can still maintain between-group variation.

Is altruism defined consistently?

In addition to ambiguity over whether CGS results in altruistic behaviour, the literature also contains multiple definitions of key terms such as altruism (Deb and Smith Reference Deb, Smith, Shackelford and Weekes-Shackelford2019). One key terminological difference already discussed is between KS-altruism and ML-altruism (either weak or strong; Figure 1). When discussing potentially altruistic behaviour in humans, some authors define altruism from a kin selection perspective (André and Morin Reference André and Morin2011; El Mouden et al. Reference El Mouden, André, Morin and Nettle2014; West et al. Reference West, Griffin and Gardner2007), others from an MLS perspective (Bowles and Gintis Reference Bowles and Gintis2011; Sober and Wilson Reference Sober and Wilson1998). However, other authors working on human cooperative evolution define altruism as short-term costs (Fehr and Fischbacher Reference Fehr and Fischbacher2003; Gintis Reference Gintis2000), even if in the long-term it may increase lifetime biological fitness, such as via reciprocity, reputation or success in group conflict.

These conflicting definitions of altruism lead to different assessments of whether cooperation is altruistic or not. For instance, MLS and short-term perspectives define altruism more broadly than kin selection. This may be one of the reasons why many authors who adopt either of these perspectives emphasise that altruism among non-relatives is rife in humans (Bowles and Gintis Reference Bowles and Gintis2011; Henrich et al. Reference Henrich, Boyd, Bowles, Camerer, Fehr, Gintis, McElreath, Alvard, Barr, Ensminger, Smith Henrich, Hill, Gil-White, Gurven, Marlowe, Patton and Tracer2005; Sober and Wilson Reference Sober and Wilson1998), while authors who adopt a kin selection view claim that cooperation towards non-kin is mutually beneficial, not altruistic (West et al. Reference West, El Mouden and Gardner2011, Reference West, Griffin and Gardner2007). Some of the controversy surrounding CGS and human cooperation may be due to different definitions of ‘altruism’, with much of the CGS literature adopting altruism as defined by MLS or short-term costs (e.g. Bowles and Gintis Reference Bowles and Gintis2011; Henrich Reference Henrich2004a; Sober and Wilson Reference Sober and Wilson1998), and other evolutionary researchers adopting the KS-altruism definition. However, as discussed above, in more recent CGS publications there has been a shift away from interpreting behaviour as altruistic, as punishment makes otherwise-costly cooperation self-interested (Boyd Reference Boyd2017). As with the differing expectations of behaviour based on maladaptive, normative and mechanism-neutral CGS described above, multiple definitions of terms such as ‘altruism’ can make it difficult to explicitly operationalise and test predictions of CGS.

Is social learning naïve?

A central premise of maladaptive CGS models is that humans naively apply cultural learning strategies, such as conformism or prestige-biased transmission, to acquire behaviour, with little concern for individual pay-offs (Boyd and Richerson Reference Boyd and Richerson1985; Henrich Reference Henrich2004a; Richerson and Boyd Reference Richerson and Boyd2005). These mechanisms are ‘blind’, in that individuals use these strategies even if they lead to maladaptive behaviour. While this may occur in some instances – evolved adaptations are never completely infallible – if these maladaptive social learning strategies persisted for long enough then it is likely that they would be outcompeted by other more sophisticated strategies (El Mouden et al. Reference El Mouden, André, Morin and Nettle2014). Relying on maladaptive social learning mechanisms to explain human cooperative behaviour is equivalent to a ‘mismatch’ between evolved cognitive mechanisms in ancestral vs modern environments; strategies such as conformism and prestige-bias may have been adaptive in our evolutionary past, but misfire and result in maladaptive behaviour in novel situations (André and Morin Reference André and Morin2011; Mace Reference Mace, Frey, Störmer and Willführ2011). Alternatively, these occasionally maladaptive strategies may persist owing to cognitive constraints. While a strategy may be adaptive on average, it may misfire in certain situations (such as regarding cooperation), and the costs to investing in a more discriminating strategy – say, a ‘conform for foraging techniques, but not cooperation’ rule – is not offset by the gains (Boyd and Richerson Reference Boyd and Richerson1985). However, as detailed below, human social learning strategies are incredibly variable and context-specific, and often give rise to adaptive behaviour, even in novel environments, giving less plausibility to both mismatch and constraint-based arguments.

To be adaptive, social learning must be strategic and context-specific (André and Morin Reference André and Morin2011; Cronk Reference Cronk2017; Laland Reference Laland2004). Non-human animals flexibly use complex social learning strategies, including ‘when’ strategies, such as ‘copy when uncertain’ or ‘copy when asocial learning is costly’, and often use multiple social learning strategies in conjunction (Laland Reference Laland2004). These strategies can lead to adaptive behaviour, such as stickleback fish using flexible pay-off-based social learning to optimally exploit resources (Kendal et al. Reference Kendal, Rendell, Pike and Laland2009), or great tits using a combination of conformism and individual reinforcement when learning to acquire adaptive foraging techniques (Aplin et al. Reference Aplin, Sheldon and McElreath2017). Given our reliance on culture one would expect the social learning strategies employed by humans to be even more discerning and context-dependent (Cronk Reference Cronk2017). Indeed, humans deploy several different adaptive social learning strategies depending on the situation (Morgan et al. Reference Morgan, Rendell, Ehn, Hoppitt and Laland2012). In economic games with the opportunity for social learning, individuals tend to use these strategies flexibly in order to maximise their earnings; that is, social learning was motivated by income maximisation, such as adopting a ‘copy-the-successful’ strategy, which in turn often reduced cooperation as the economically successful individuals were the most selfish (Burton-Chellew et al. Reference Burton-Chellew, Nax and West2015, Reference Burton-Chellew, El Mouden and West2017). Although experiments have shown that individuals do use frequency-based information in social dilemmas, this is distinct from conformism as it is not neutral with regard to pay-offs and does not reduce variation within groups (van den Berg et al. Reference van den Berg, Molleman and Weissing2015). Thus, cultural learning mechanisms which stabilise and reduce group differences, such as conformism, are not frequently observed in experiments, especially in cooperative situations (Efferson et al. Reference Efferson, Richerson, McElreath, Lubell, Edsten, Waring, Paciotti and Baum2007; Eriksson and Coultas Reference Eriksson and Coultas2009; Lamba Reference Lamba2014). Laland has summarised this succinctly; ‘Human beings copy; they copy a great deal. But they do not slavishly copy. Slavish copying would not be adaptive’ (Laland Reference Laland2017, p.64). This is not to say that culturally acquired behaviour will always be adaptive. Clearly social learning rules can be misapplied and result in inefficient or maladaptive behaviour (Laland and Williams Reference Laland and Williams1998; Richerson and Boyd Reference Richerson and Boyd2005), but where cultural traits impact biological fitness and culture causes maladaptive behaviour, there will be selection pressure in favour of alternative strategies which enhance biological fitness (El Mouden et al. Reference El Mouden, André, Morin and Nettle2014).

Do humans learn cooperative behaviour culturally?

For CGS to occur, cultural mechanisms must stabilise and reduce within-group variation in behaviour. Several aspects of human behaviour are indisputably cultural as they could not have originated within one generation; think complex technology (Henrich Reference Henrich2004b), institutional complexity (Currie et al. Reference Currie, Greenhill, Gray, Hasegawa and Mace2010) and religion (Watts et al. Reference Watts, Greenhill, Atkinson, Currie, Bulbulia, Gray and Watts2015). Problems occur, however, in applying the same logic to individual-level behaviours such as cooperation, which can emerge via several pathways, including fixed genetic behaviours, phenotypic plasticity, epigenetic inheritance, individual learning and social learning (or combinations thereof; Jablonka and Lamb Reference Jablonka and Lamb2014). Unlike complex technology which is absent outside humans, many other species – including bacteria (Kümmerli et al. Reference Kümmerli, Jiricny, Clarke, West and Griffin2009) – display variability and plasticity in cooperative behaviour (Adams et al. Reference Adams, Robinson, Mannarelli, Hatchwell and Adams2015). As cooperation can evolve in the absence of social learning, this means that there are multiple proximate pathways by which cooperative behaviour can evolve. Many of these make similar predictions regarding the distribution of cooperative behaviour, making them difficult to differentiate. For instance, both phenotypic plasticity and conformism predict that, given different environments, behaviour within groups will be more similar than behaviour between groups, irrespective of genetic differences.

It is often assumed that differences in cooperative behaviour are socially learned, such as by conformism (Boyd and Richerson Reference Boyd and Richerson1985), copying cooperative prestigious individuals (Henrich et al. Reference Henrich, Chudek and Boyd2015) or pay-off-biased imitation (Boyd and Richerson Reference Boyd and Richerson2002). While individuals do appear to learn some cooperative behaviour via social learning, when it is employed it tends to be based on pay-off-biased transmission (Burton-Chellew et al. Reference Burton-Chellew, El Mouden and West2017), rather than conformism (Lamba Reference Lamba2014; little empirical work has explored whether adults learn cooperative behaviour via prestige-bias; see Henrich et al. Reference Henrich, Chudek and Boyd2015). Individuals also readily adapt their behaviour via individual evaluations of the situation: for instance, in the absence of social learning individuals alter their cooperative behaviour in consistently adaptive directions according to cues of anonymity (Ernest-Jones et al. Reference Ernest-Jones, Nettle and Bateson2011), whether competition is within or between groups (Puurtinen and Mappes Reference Puurtinen and Mappes2009), and whether interactions are repeated (Rand and Nowak Reference Rand and Nowak2013). Furthermore, individuals vary their evaluations of resource distributions based on local ecological factors such as the role of luck in acquiring resources, social homogeneity, warfare and resource availability (Nettle and Saxe Reference Nettle and Saxe2019). This variation can be explained by individual-level adaptive decision-making, without the need for acquiring cooperative behaviour culturally. This provides an alternative explanation for cross-cultural differences in cooperation (e.g. Henrich et al. Reference Henrich, Boyd, Bowles, Camerer, Fehr, Gintis, McElreath, Alvard, Barr, Ensminger, Smith Henrich, Hill, Gil-White, Gurven, Marlowe, Patton and Tracer2005), without ascribing these differences to cultural learning. This is not to say that cultural transmission is unimportant for acquiring cooperative behaviour, but at present this is largely an untested assumption and probably occurs in combination with other processes (Lamba Reference Lamba2014). However, the evidence for widespread conformism – a key mechanism for stabilising group behaviour – is rather weak.

Do norms determine behaviour?

Normative CGS proposes that individuals acquire arbitrary group norms of cooperation owing to a norm psychology, where individuals adopt the norms of the local group and punish norm-violators. Norms are ‘learned behavioural standards shared […] by a community’ (Chudek and Henrich Reference Chudek and Henrich2011, p. 218). It is important to differentiate this normative definition of ‘norm’ from descriptions of actual behaviour – that is, distinguishing between how individuals ought to act and how they actually do act (Wallen and Romulo Reference Wallen and Romulo2017). As norms are by definition group-level traits, one would expect low levels of within-group variation in stated norms, beliefs and values, as appears to be the case (Bell et al. Reference Bell, Richerson and McElreath2009; although see Eriksson and Coultas Reference Eriksson and Coultas2009). However, as actual behaviour, not norms or beliefs, drives evolution, the extent to which norms translate into behaviour is a key question.

Normative CGS requires that cultural norms cause cooperation (Figure 4, upper). In some cases the association between norms and behaviour is strong, such as Maasai osotua needs-based cooperation (Cronk Reference Cronk2007), food-sharing among Samburu pastoralists (Lesorogol Reference Lesorogol2007) and reciprocity among Saami reindeer herders (Thomas et al. Reference Thomas, Bårdsen and Næss2018). Elsewhere, there are discrepancies between norms and behaviour. As predicted by genetic evolutionary models, Mukogodo parents preferentially invest in daughters over sons as they live in relative poverty compared with their neighbours, so males are less able to compete in the mating market. However, there is a mismatch as the stated Mukogodo norm is to prefer sons, a norm probably borrowed from their higher-status Maa-speaking neighbours (Cronk Reference Cronk2017). Tibetan pastoralists display the same mismatch: norms which prefer sons, yet daughter-biased parental investment is observed (Du and Mace Reference Du and Mace2017). If the norms individuals espouse are divorced from behaviour, then norms cannot always shape individual actions.

Figure 4. Determinants of cooperative behaviour from a cultural group selection ‘norm first’ perspective (upper), a socioecological ‘ecology first’ approach (middle) and a combined approach (bottom). ‘Socioecology’ is defined as the social, economic and physical environment, so includes subsistence patterns, demography/group size, group competition, sedentarisation, etc., as well as culturally evolved behaviours (e.g. residence patterns, technology). ‘Behaviour’ is how individuals actually behave, ‘norms’ are shared beliefs about the ‘correct’ behaviour in a group, while ‘institutions’ are structures which shape social interactions and alter the pay-offs to cooperation and defection. Cultural group selection (upper): cultural group selection is generally silent about the role of socioecology and tends to portray cooperative behaviour as solely determined by norms. Socioecological approach (middle): socioecology (broadly defined) determines behaviour, which in turn may influence shared norms (the shaded arrow from behaviour to norms). Combined approach (bottom): both the socioecology and culturally evolved norms/institutions impact behaviour, with behaviour also evaluating and updating existing norms/institutions (where possible) and shaping socioecological circumstances (e.g. by niche construction). Norms and institutions therefore also impact socioecology in path-dependent ways, while current socioecological circumstances – which may be norm/institution-dependent – shape and constrain future norms and institutions. Thus, norms/institutions and socioecology feed into one another and may be difficult to separate in practice (see Box 2). This combined approach involving reciprocal feedback loops is likely to be necessary in explaining large-scale cooperation.

Furthermore, a correlation between norms and cooperation is not necessarily evidence that norms cause behaviour. In many cases behaviour is shaped by the socioecology, which in turn shapes norms (Figure 4, middle). For instance, among Chinese farmers historical ecological differences in subsistence practices (rice- vs wheat-farming) are associated with present-day behavioural and psychological differences, as rice-farming requires greater cooperation and interdependence than wheat-farming (Talhelm et al. Reference Talhelm, Zhang, Oishi, Shimin, Duan, Lan and Kitayama2014; although for a re-evaluation of this claim see Roberts Reference Roberts2015). Similarly, within-society variation in cooperation – where society-level norms are relatively constant – can be explained by differences in socioecology, such as demography, need and social networks (Lamba and Mace Reference Lamba and Mace2011; Smith et al. Reference Smith, Dyble, Thompson, Major, Page, Chaudhary, Salali, Vinicius, Migliano and Mace2016). This ‘ecology first’ approach (as opposed to the CGS ‘norm first’ approach) can also explain cases where behaviour may be adaptive but the norms are outdated (Du and Mace Reference Du and Mace2017), owing to cultural lag.

Why follow norms and punish norm-breakers?

Behaviour determining norms cannot explain all aspects of human cooperation. Many examples presented above are individual choices (such as sex-biased parental investment) or small-scale cooperation (such as forager food-sharing) which plausibly do not require norms in order to evolve (although these behaviours may be enshrined in norms after their original evolution). Large-scale cooperative ventures, such as warfare and common pool resource management (resource extraction, irrigation systems, fishing rights, etc.) are more difficult to evolve, given that the scope for cooperation declines with group size (Powers and Lehmann Reference Powers and Lehmann2017). CGS may provide a plausible explanation for these large-scale cooperative ventures.

While maladaptive CGS can lead to altruistic behaviour, normative CGS is generally considered consistent with self-interest. That is, when cultural norms are backed up by punishment of norm-violators, it is in an individual's self-interest to cooperate and follow group norms (Boyd Reference Boyd2017; Boyd and Richerson Reference Boyd and Richerson2009a; Chudek and Henrich Reference Chudek and Henrich2011; Zefferman and Mathew Reference Zefferman and Mathew2015). The famous ‘second-order free-rider problem’ comes into play here; even though everyone is better off by cooperating and punishing free-riders, individuals who free-ride on others’ altruistic punishment will have greater fitness than punishers, causing the population to evolve towards non-cooperation. In small groups or situations of strong between-group competition, this second-order free-rider problem may not arise as individuals may increase their direct fitness by punishing, even though others in the same group will have higher fitness (assuming groups are composed of non-kin, in these cases the behaviour would be weakly ML-altruistic, not strongly ML-altruistic or KS-altruistic; Lehmann et al. Reference Lehmann, Rousset, Roze and Keller2007). However, in larger groups individuals are unlikely to gain direct fitness benefits by cooperating, even with group competition present. This occurs because in large groups punishment is unlikely to have a significant impact on group success, so the direct fitness gains are minimal and unlikely to outweigh the costs (Boyd Reference Boyd2017; Powers and Lehmann Reference Powers and Lehmann2017).

A potential solution to this problem is that cooperation can be stable if punishing norm-violators is itself normative (Boyd Reference Boyd2017; Boyd and Richerson Reference Boyd and Richerson1992). That is, individuals punish norm-breakers simply because they are violating a norm. Competition between groups can then act to select the most cooperative equilibria. The question then becomes: why is punishing norm-violators normative? One proposed answer is based on an evolved norm psychology which arose in the context of small groups where there were direct benefits to punishing norm-breakers. With this psychological machinery in place, it could then be harnessed via CGS for large-scale cooperative projects (Boyd Reference Boyd2017, pp. 118–120).

Normative CGS provides a plausible theory for large-scale human cooperation and makes several predictions which can be tested empirically, some of which are not unique to CGS (e.g. reputation-based theories also predict that humans should take an interest in third-party behaviour). One unique prediction is that individuals should indiscriminately punish norm-breakers simply because they broke a norm, and not for any direct fitness benefit (other than to avoid potential higher-order punishment for not punishing).

Despite theoretical coherence, the evidence that individuals indiscriminately punish norm-violators is rather weak. Some experiments have shown that individuals punish others at a cost to self, even in one-shot interactions (Bernhard et al. Reference Bernhard, Fischbacher and Fehr2006; Fehr and Fischbacher Reference Fehr and Fischbacher2004; Henrich et al. Reference Henrich, Boyd, Bowles, Camerer, Fehr, Gintis, McElreath, Alvard, Barr, Ensminger, Smith Henrich, Hill, Gil-White, Gurven, Marlowe, Patton and Tracer2005). This behaviour is prima facie support for a norm psychology, and is often claimed to be altruistic (Bowles and Gintis Reference Bowles and Gintis2011; Boyd et al. Reference Boyd, Gintis, Bowles and Richerson2003). However, in repeated situations these seemingly altruistic individuals increase levels of cooperation, often resulting in greater pay-offs for the group and for themselves (dos Santos et al. Reference dos Santos, Rankin and Wedekind2013; Gürerk et al. Reference Gürerk, Irlenbusch and Rockenbach2006). These ‘altruistic punishers’ also receive other direct benefits, such as a reputation for trustworthiness (Jordan et al. Reference Jordan, Hoffman, Bloom and Rand2016) or other reputational or monetary incentives (Ostrom Reference Ostrom1990). Experiments which have pitted predictions of normative CGS (in which punitive behaviours are indiscriminately directed towards norm-violators) against predictions from a self-interested model (in which punishment brings greater future cooperation to the punisher) have found that behaviour conformed to the latter. This suggests that individuals do not punish defectors solely to maintain the group norm, but rather punish selectively to encourage cooperation and increase future pay-offs for themselves (Krasnow et al. Reference Krasnow, Cosmides, Pedersen and Tooby2012; see also Burton-Chellew and West Reference Burton-Chellew and West2013; Pedersen et al. Reference Pedersen, McAuliffe and McCullough2018). Thus, while punishment of norm-breakers certainly does occur (Mathew and Boyd Reference Mathew and Boyd2011; Wiessner Reference Wiessner2005), in contrast to predictions made by normative CGS, this punishment appears self-interested (beyond avoiding higher-level punishment).

Do group differences arise via stochastic processes and are norms stable?

For CGS to occur, there must be differences between groups. In maladaptive CGS the origin of groups at different cooperative equilibria is not formally modelled, but rather is assumed to occur via stochastic processes (Boyd et al. Reference Boyd, Richerson and Henrich2011). In normative CGS, groups vary randomly as a result of punishment and reward processes which reach different stable equilibria; some groups develop cooperative norms, others do not (Boyd Reference Boyd2017; Ensminger and Henrich Reference Ensminger, Henrich, Jean and Henrich2014). As stated by Boyd and Richerson (Reference Boyd and Richerson2009a, p. 3273) ‘[t]he problem is that the three Rs [reputation, reciprocation and retribution] can stabilise any behaviour. If everybody agrees that individuals must do X, and punish those who do not do X, then X will be evolutionarily stable as long as the costs of being punished exceed the costs of doing X. It is irrelevant whether X benefits the group or is socially destructive. It will pay to do X’. Although maladaptive and normative CGS processes posit different proximate mechanisms, they agree that different groups will reach different equilibria, after which equilibrium selection can act.

While these are plausible mechanisms to stabilise group-level behaviours, theories of institutional change based on individuals or groups creating cooperative institutions provide an alternative explanation which may be better supported by empirical evidence. Maladaptive and normative CGS predict that behaviour within groups is stable, so even if superior equilibria exist it will be difficult for groups to change their norms endogenously. However, studies on institutional change indicate that individuals are often aware that they are in a poorly functioning group and frequently try and create new rules to enhance fitness (Ostrom Reference Ostrom1990; Singh et al. Reference Singh, Wrangham and Glowacki2017). For instance, in the 1970s the Turkish fishing village of Alanya was suffering from an over-exploitation of fishing stocks. After a decade of trial-and-error, the village settled on a new system for governing fishing rights which benefits everyone and permits cheaters to be detected and punished swiftly (Ostrom Reference Ostrom1990, pp. 18–20). This institutional change arose based on rational self-interest, without requiring conformism/prestige-bias, a norm psychology or group competition. The change was also endogenous, rather than copying the institutions of more successful groups, so cannot be attributed to CGS.

‘Rational-actor’ theories of institutional change can explain how individuals cooperate in large-scale projects via self-interest, without relying on altruistic cooperation, maladaptive social learning strategies or a norm psychology (Powers et al. Reference Powers, van Schaik and Lehmann2019). One way institutions solve the second-order free-rider problem is by rewarding punishment, thereby aligning punishment with self-interest. Examples include punishers keeping the fines taken from defectors, as found in many common-pool resource systems worldwide (Ostrom Reference Ostrom1990). If cooperative reputations are known and the exclusion of individuals with poor reputations is possible, then institutions which ostracise defectors from partaking in subsequent mutually beneficial interactions can evolve which also avoid the second-order free-rider problem (Greif Reference Greif1993; Panchanathan and Boyd Reference Panchanathan and Boyd2004). Each of these processes results in the within-group selection term in the multi-level Price equation being positive, meaning that cooperators have greater fitness than defectors within groups (or, from a kin selection view, the individuals increase their direct fitness by cooperating). Put another way, individual-level processes – not just competition between groups – can lead to the spread of cooperative behaviour (Chaudhary et al. Reference Chaudhary, Salali, Thompson, Rey, Gerbault, Stevenson, Dyble, Page, Smith, Mace, Vinicius and Migliano2016). Formal models of voluntary investment in group-beneficial institutions have indicated that this process can result in stable systems of costly institutions and can evolve from an asocial state (Powers and Lehmann Reference Powers and Lehmann2013). This also means that that stochastic processes are not required to explain the origins of between-group variation, as required by models of maladaptive and normative CGS.

This institution-based account is not necessarily a wholesale alternative to CGS. If groups are in competition then both CGS and institutional accounts predict that more-cooperative groups will prosper at the expense of less-cooperative groups (Richerson et al. Reference Richerson, Baldini, Bell, Demps, Frost, Hillis, Mathew, Newton, Naar, Newson, Ross, Smaldino, Waring and Zefferman2016). The difference between these theories occurs in the mechanisms which give rise to between-group variation in cooperative behaviour. Normative CGS relies on a norm psychology to punish norm-violators, while maladaptive CGS relies on social learning mechanisms to stabilise group behaviour. Although the specific psychological mechanisms are not modelled, institutional approaches rely on individuals negotiating rules to promote cooperation, coordinate behaviour, punish defectors and reward those who punish free-riders. In this institutional view the second- (or third-, or fourth-, or nth-) order free-rider problem is not solved by the punishment of norm-violators being normative or by content-blind social learning mechanisms, but rather by rewards, monitoring mechanisms and incentives to coordinate behaviour and align punishing with self-interest.

Institutional accounts also diverge with predictions from CGS in another respect. In CGS, behaviour is stable within groups, meaning that group competition is necessary to select the more cooperative groups (Boyd Reference Boyd2017; Ensminger and Henrich Reference Ensminger, Henrich, Jean and Henrich2014). In contrast, institutional accounts do not necessarily require group competition in order for groups to become more cooperative, as this process can occur endogenously (Powers and Lehmann Reference Powers and Lehmann2013; although in other circumstances group competition may be required to uphold otherwise-costly institutions, see Turchin et al. Reference Turchin, Currie, Turner and Gavrilets2013). Group competition may of course facilitate the spread of group-beneficial institutions, but group conflict is not a necessary condition for these institutions to evolve (in this respect, when considering group competition institutional accounts can be thought of as examples of mechanism-neutral CGS: see Table 1).

From cross-sectional data it is difficult to distinguish between these alternative theories, as all predict that group-beneficial behaviour should be relatively common. Longitudinal and historical data are important sources required to determine the processes which lead to the emergence of institutions. While current evidence is rather fragmentary, several case studies, such as fishing rights in Alanya, demonstrate how individuals negotiate group-beneficial norms and institutions within groups (Ostrom Reference Ostrom1990; Singh et al. Reference Singh, Wrangham and Glowacki2017). Ostrom provides a detailed example of institutional change in sustainable groundwater basin extraction in California which arose endogenously and without group competition (see Ostrom Reference Ostrom1990, chapter 4; although surrounding districts later borrowed and adapted these institutions, which could be considered CGS). Experiments have also demonstrated how communication and punishment can lead to the emergence of cooperative institutions in the absence of group conflict (Ostrom Reference Ostrom2006). Additionally, although many cultural traits possess a phylogenetic signal, suggesting long-term cultural stability (Mathew and Perreault Reference Mathew and Perreault2015), other traits relating to cooperation, such as food-sharing, show little phylogenetic signal (Ringen et al. Reference Ringen, Duda and Jaeggi2019). Studies on forager cooperation have shown that cooperative behaviour is variable within ethnographic groups (Smith et al. Reference Smith, Larroucau, Mabulla and Apicella2018), often based on differences in ecology (Lamba and Mace Reference Lamba and Mace2011; Smith et al. Reference Smith, Dyble, Thompson, Major, Page, Chaudhary, Salali, Vinicius, Migliano and Mace2016), and changeable within the space of a single generation (Gurven et al. Reference Gurven, Hill and Kaplan2002). These findings are consistent with this lack of phylogenetic signal and suggest that cooperative behaviour can display a rapid adaptive response to changing environments. This rapid endogenous adaptive change is difficult to reconcile with CGS, as the time taken for cultural traits to spread via inter-group conflict is likely to be upwards of 500 years (Soltis et al. Reference Soltis, Boyd and Richerson1995: although other CGS processes of selective migration and copying successful groups can occur on faster time-scales, they are unlikely to explain these patterns of short-term adaptive variation).

A further prediction of CGS which can be tested is norm content. Maladaptive and normative CGS predict that mechanisms to stabilise behaviour should lead to groups at multiple different equilibria, many of which include neutral or positively harmful traits (Ensminger and Henrich Reference Ensminger, Henrich, Jean and Henrich2014). Many cultural traits are certainly puzzling from a purely biological fitness-maximising perspective (Edgerton Reference Edgerton1992), but whether they actually damage fitness is harder to assess. Many seemingly harmful cultural practices may have an adaptive basis, such as lethal warfare (Chagnon Reference Chagnon1988), female genital cutting (Howard and Gibson Reference Howard and Gibson2017) and witchcraft accusations (Mace et al. Reference Mace, Thomas, Wu, He, Ji and Tao2018), while for other traits if their effects on fitness are negligible then they could have plausibly evolved by cultural drift (El Mouden et al. Reference El Mouden, André, Morin and Nettle2014). In a survey of 60 cultures, the behaviours considered ‘moral’ were strikingly uniform across all societies and are predicted by evolutionary models (Curry et al. Reference Curry, Mullins and Whitehouse2019). These include helping kin, helping the group, reciprocity and bravery. This lack of cross-cultural variation argues against the CGS idea that norms can stabilise any behaviour. Rather, the behaviours which are normative tend to be broadly fitness-enhancing. Of course, perhaps all of the poorly adapted groups with anti-social norms have already been replaced, but the studies above detailing how new norms and institutions arise suggest that they are generally tailored to be both group- and individually beneficial, rather than owing primarily to stochastic processes.

While additional data is required, the current evidence argues against CGS, which suggests that group-level differences in cooperation are largely based on social learning strategies and/or a norm psychology to stabilise any cooperative behaviour. Instead, based on the current technological, social, ecological, institutional and economic constraints, individuals and groups try to design norms and institutions which benefit themselves; as individual fitness is frequently tied to group fitness, norms and institutions are likely to be broadly group-beneficial. Norms and institutions are continually being evaluated, contested and (re-)shaped (Morin Reference Morin2016a; Ostrom Reference Ostrom1990; Singh et al. Reference Singh, Wrangham and Glowacki2017), arguing against the stability of norms. There will still probably be selection between different norms/institutions, as some will be better-suited than others in a given ecology, but this process does not necessarily require norm-compliance and randomly formed group-level stable strategies, as proposed by maladaptive and normative theories of CGS.

Does cultural group selection act on cultural fitness, biological fitness, or both?

Although CGS may impact genetic evolution and biological fitness, primarily it is concerned with the spread of cultural traits. As stated by Richerson et al. (Reference Richerson, Baldini, Bell, Demps, Frost, Hillis, Mathew, Newton, Naar, Newson, Ross, Smaldino, Waring and Zefferman2016, p. 5), ‘[a] cultural variant that attracts many imitators has a reproductive success which is not necessarily tied to biological reproduction. [In the case of CGS] we can use models to specify the costs, benefits, and success of cultural variants in terms of changes in their frequency’. This cultural fitness perspective can be expressed using a cultural multi-level Price equation (see Section S2 of the Supplementary Information). However, a focus solely on cultural fitness does not necessarily equate to an impact on biological fitness.

A key distinction is between cultural traits which impact biological fitness (reproductive success or inclusive fitness) and cultural traits which impact cultural fitness (the cultural influence of an individual/group for a specific trait). Birch (Reference Birch2017) refers to these as Cultural Selection 1 (CS1) and Cultural Selection 2 (CS2), respectively. CGS always acts on CS2 (the cultural trait's impact on cultural fitness) but does not always act on CS1 (the cultural trait's impact on biological fitness). Take competition between economic organisations (Richerson et al. Reference Richerson, Baldini, Bell, Demps, Frost, Hillis, Mathew, Newton, Naar, Newson, Ross, Smaldino, Waring and Zefferman2016). Certain businesses have greater cultural fitness than others, in terms of either survival or being imitated. This can be thought of as CGS acting via cultural fitness (CS2). However, the impact of being part of a successful business on biological fitness (CS1) may be minimal, given that employees from unsuccessful organisations are unlikely to have lower reproductive success (they can just get jobs elsewhere). In this example CS1 and CS2 are decoupled, so although more successful businesses have greater cultural fitness, this has little impact on biological fitness.

In other situations, cultural and biological fitness are more tightly linked. Warfare may be a good example. Success in warfare frequently depends on cultural factors such as social organisation and technological innovations. Groups possessing these traits are able to out-compete less-organised and less technologically complex groups (Diamond Reference Diamond1997; Turchin et al. Reference Turchin, Currie, Turner and Gavrilets2013), meaning that being part of such groups may enhance both biological and cultural fitness. However, vanquished individuals from losing groups are often assimilated within the winning group, rather than killed outright (Richerson and Boyd Reference Richerson and Boyd2005; Soltis et al. Reference Soltis, Boyd and Richerson1995). Cultural and biological fitness will be completely aligned if individuals from losing groups are eliminated, while if warfare is largely symbolic with few casualties and individuals from losing groups assimilated within the victorious group, then selection acts predominantly on cultural fitness, not biological fitness (the cultural spread of cooperative groups may favour subsequent selection on genes for cooperativeness – see grey arrow in Figure 3 – but this selection is distal, rather than immediate).

In practice it is often difficult to distinguish between these processes. Consider the Nuer/Dinka conflict – frequently presented as a canonical case of CGS (Richerson and Boyd Reference Richerson and Boyd2005) – where the superior organisation of the Nuer in inter-group conflict led to them expanding at the Dinka's expense. Although some Dinka were killed, many Dinka assimilated into the Nuer population. A similar process combining both replacement and assimilation has probably occurred throughout human history. Although these processes appear similar when adopting a cultural fitness perspective (CS2) – as the focus is on the spread of the cultural trait – they are different when adopting a biological fitness perspective (CS1); group expansion and conquest without assimilation mean that losing individuals have lower biological fitness, while assimilation does not necessarily entail a decrease in biological fitness to individuals from the losing group. The distinction between cultural traits acting on cultural fitness (CS2) and cultural traits acting on biological fitness (CS1) may be a useful framework for thinking about CGS and whether one is interested in the spread of cultural traits themselves or on the immediate impact of these cultural traits on biological fitness.

Is cultural group selection a group-level process?

Related to the point above, CGS proposes several mechanisms by which group-beneficial traits can spread (Henrich Reference Henrich2004a, Reference Henrich2015; Richerson et al. Reference Richerson, Baldini, Bell, Demps, Frost, Hillis, Mathew, Newton, Naar, Newson, Ross, Smaldino, Waring and Zefferman2016; Richerson and Boyd Reference Richerson and Boyd2005), which are summarised in Figure 5. Some of these mechanisms are group-level processes which act on populations, such as one group outcompeting another, either directly (via warfare or conquest) or indirectly (via higher reproductive rates, leading to faster diffusion rates into unoccupied environments). As discussed above, in these processes biological fitness (CS1) and cultural fitness (CS2) may be aligned, in that more cooperative and organised groups possess greater cultural and biological fitness (although at other times group competition may spread just cultural traits, with little impact on biological fitness). These are mechanisms of natural selection – either cultural or biological – acting on groups. However, other proposed CGS processes are decision-making mechanisms of cultural transmission. One of these processes is individuals copying successful traits of other groups (Boyd and Richerson Reference Boyd and Richerson2002), while another is selective migration to successful groups (Boyd and Richerson Reference Boyd and Richerson2009b). While this could be viewed as cultural selection – as some traits are preferentially imitated over others, meaning they have higher cultural fitness – from the perspective of biological fitness no selection is occurring.

Figure 5. Different cultural group selection processes which result in the spread of group-beneficial norms and behaviours. Angel icons represent cooperators and devil icons represent defectors. (a) Differential expansion rates: in the absence of conflict, cooperative groups may expand into unoccupied territory at a faster rate than less-cooperative groups. (b) Inter-group conflict: if there is competition between groups, then more-cooperative and cohesive groups will expand at the expense of less-cooperative groups. (c) Copying successful groups: individuals from less-successful groups copy the cooperative traits of successful groups. (d) Differential migration: individuals from less-successful groups migrate to more-successful groups and adopt their traits. From a biological fitness perspective, processes (a) and (b) may be examples of a selection process acting on groups, while in processes (c) and (d) no selection on biological fitness is occurring as these are choices made by individuals or groups to ostensibly increase future biological fitness. From a cultural fitness perspective, all could be group selection processes acting on cultural fitness (as cooperative groups are more likely to propagate).

Depending on whether one adopts a cultural fitness or biological fitness perspective, these processes have different interpretations. In terms of cultural fitness, copying traits from successful groups and selective migration may be group-level processes (and hence CGS), as groups possessing certain cultural traits are more likely to be imitated. However, from a biological fitness perspective, these processes are choices by individuals or groups ostensibly attempting to maximise their future biological fitness. While both of these demonstrate how beneficial cultural traits can spread, there is no need to invoke group-level selection processes acting on biological fitness here as individuals (or groups) either adopt, or migrate then adopt, the practices of more successful groups (Amir et al. Reference Amir, Jordan and Rand2016; Morin Reference Morin2016b).

Is cultural group selection an alternative to approaches based on biological fitness?

Some proponents of CGS appear to view it as an alternative to existing approaches based on biological fitness. In a recent review, Richerson et al. (Reference Richerson, Baldini, Bell, Demps, Frost, Hillis, Mathew, Newton, Naar, Newson, Ross, Smaldino, Waring and Zefferman2016, p. 3) claim that ‘[t]he issue here is not whether effects such as nepotism, reputation building, and other mechanisms of cooperation supported by reciprocity and inclusive fitness exist […] but whether they are sufficient to explain the large-scale cooperation in human societies’. However, when the focus is on biological (rather than cultural) fitness, cultural and biological approaches are difficult to separate, as cultural adaptations often work via mechanisms of reputation, reciprocity, kinship and punishment to enhance biological fitness (Henrich and Boyd Reference Henrich and Boyd2016; Henrich and Henrich Reference Henrich and Henrich2007). Witchcraft accusations (Mace et al. Reference Mace, Thomas, Wu, He, Ji and Tao2018) and storytelling (Smith et al. Reference Smith, Schlaepfer, Major, Dyble, Page, Thompson, Chaudhary, Salali, Mace, Astete, Ngales, Vinicius and Migliano2017) are clear examples of cultural factors which impact reputation and shape patterns of cooperation. Kinship norms may also be culturally evolved institutions to coordinate behaviour and facilitate cooperation (Cronk et al. Reference Cronk, Steklis, Steklis, van den Akker and Aktipis2019). For instance, Lamaleran kinship is based on patrilineages which coordinate collaborative whaling ventures by determining who can work together, thereby avoiding conflict (Nolin Reference Nolin2011). In online marketplaces, institutions facilitate large-scale cooperation via reviewer systems to broadcast sellers’ (and buyers’) reputation, upon which individuals can base future mutually beneficial interactions (Livingston Reference Livingston2005).

While these norms and institutions are certainly products of cultural evolution, whether they evolved via CGS (as opposed to arising endogenously) is unclear. Even if some did evolve by CGS (for instance, online marketplace companies with reviewer systems outcompeting companies without reviewer systems), this is not necessarily in conflict with traditional evolutionary explanations, as these successful institutions harness these kinship, reciprocal, reputational or sanctioning mechanisms more effectively to facilitate large-scale cooperation (Henrich and Boyd Reference Henrich and Boyd2016). If CGS leads to maladaptive or altruistic behaviour – as claimed by some (Bowles and Gintis Reference Bowles and Gintis2011; Richerson and Boyd Reference Richerson and Boyd2005) – then it would be an alternative to these traditional evolutionary explanations. However, much culturally mediated cooperation appears broadly consistent with biological self-interest. Rather, CGS provides a potential answer to how different societies arrive at different norms and institutions which harness these traditional evolutionary explanations for large-scale cooperation. However, knowing that CGS occurred tells one nothing about the mechanism(s) by which one group out-competed another. When considering biological fitness, CGS is simply multi-level selection with a focus on culture as a proximate mechanism by which individuals and groups acquire adaptive behaviour (for additional discussion of the status of culture as a proximate mechanism, see Box 2). CGS does diverge from traditional evolutionary accounts when the focus is on cultural, rather than biological, fitness, but this is because traditional evolutionary accounts which focus on ultimate biological fitness outcomes tend not to be concerned with cultural fitness and the proximate psychological mechanisms underpinning decision-making and cultural transmission.