Nature © Macmillan Publishers Ltd 1998 8 A cts of giving, receiving and repaying permeate various aspects of human life and, similarly, the lives of many animals. However, evolutionary theory pre- dicts that reciprocally altruistic cooperation should be rare in the animal kingdom — natural selection should tend to favour selfish, uncooperative behaviour. So what are the special conditions required for reciprocal altruism to occur? And what is the optimal level of altruism that each partner should show? New mathematical models reported by Roberts and Sherratt 1 on page 175 of this issue show that cooperation can spread more readily than previously thought. Moreover, cooperation should increase over the history of a partnership when organisms repeatedly interact, and when they can flexibly respond in a continu- ous manner to the level of cooperation that was previously shown by their partners. Theoretical work on the evolution of altruism between unrelated individuals has been dominated by the ‘prisoner’s dilemma’ model. In an encounter, each of two individ- uals can choose either to cooperate or not to cooperate (‘defect’ in trade jargon). If they both cooperate, both do better than if they had both defected. But if one player defects while the other cooperates, the defector gets the highest reward and the cooperator gets nothing, or even suffers a net cost. Clearly, if players meet only once the best option is to defect. However, if the players meet often enough, cooperation can be favoured 2 . Computer tournaments between different strategies found that the most effective strat- egy was ‘tit-for-tat’, whereby players cooper- ate on the first move and thereafter repeat the opponent’s previous move. In this way, they can punish defection but resume coopera- tion once the opponent does 3 . Over the past 15 years there have been hundreds of theoretical papers aimed at working out which strategy would beat tit- for-tat under a range of conditions — for example, when individuals occasionally misidentify what opponents have done dur- ing the previous move 4 . Although of intrin- sic theoretical interest, these hypothetical contests have been somewhat frustrating for empiricists, who have found it difficult to document the prisoner’s dilemma array of fitness payoffs (that is, reproductive consequences for cooperating and defect- ing) in real biological systems 5,6 . In particu- lar, a biologically unrealistic limitation of the conventional prisoner’s dilemma is that it assumes that players can choose only between all-out cooperation and utter defection. But it is clear that coopera- tion is rarely an ‘all or nothing’ behaviour (Fig. 1). Roberts and Sherratt 1 allowed for vari- able investment by each partner — with zero investment corresponding to defection — and analysed what would be the best strategy to adopt. Their computer simulations showed that, among several strategies com- pared, the most effective was ‘raise the stake’ (RTS). This strategy raises its investment with individuals who have matched or bet- tered their partner’s last move. RTS was able to spread in the population until fixed when in competition with strategies which, for example, never invest, match a partner’s investment, undercut a partner, or cheat probabilistically. Under most conditions, RTS was also successful in resisting sub- sequent invasion by any of these defector strategies. The success of RTS stems from the fact that it is resistant to exploitation by defector strategies, while making the most of cooperative opportunities. In short, the success of RTS is due to the same features (being nice, retaliating and forgiving) that make tit-for-tat so successful in the discrete model of the prisoner’s dilemma. Perhaps the most important prediction to come from RTS is that individuals should invest relatively little during their first encounters with new partners, and then become increasingly cooperative with them. By ‘testing the water’, cooperators limit their losses but capitalize on cooperative opportu- nities with individuals who are prepared to reciprocate. This makes sense. Are you more likely to lend £100 to a friend or to a stranger? The idea of earlier encounters being used to build trust has also been proposed by May 7 . He suggested that partners may eventually undertake an enterprise with a large reward — say, robbing a bank — once they have built up enough trust during previous en- counters. Unfortunately, empirical studies recording changes in the degree of coopera- tion during successive interactions are scarce. Experiments with humans showed that people become increasingly cooperative during four successive interactions 8 . But, in that study, the options were only discrete cooperation or defection, as in the original prisoner’s dilemma game. Roberts and Sherratt 1 also suggest that potlatch (a custom among certain North American Indians whereby property is given away or destroyed) may represent altruistic escalation. However, some caution is needed here as potlatches may serve as honest signals NATURE | VOL 394 | 9 JULY 1998 121 news and views Familiarity breeds cooperation Laurent Keller and H. Kern Reeve Many theoretical models have been developed to study the conditions under which unrelated individuals should cooperate or not cooperate. But such behaviour is rarely ‘all or nothing’, and new mathematical models allow the optimal level of cooperation to be determined. Figure 1 You scratch my back and I’ll scratch yours — grooming by Barbary macaques on Gibraltar. Roberts and Sherratt 1 have devised new models for cooperation that take into account the fact that, in biological situations, the choice is not always as clear-cut as whether or not to cooperate. One example of this is the amount of time that macaques who cooperate by grooming one another invest in this altruistic act. R. D. MARTIN, AIM, ZURICH