Figure 1. Average payoffs across the four treatments for humans (empty bars) and the ZD strategies implemented by the computer programme (filled bars).

In line with the theory, extortioners succeed against their human co-players, whereas generous ZD strategies lag behind their human opponents. Throughout the paper, we use two-tailed non-parametric tests for our statistical analysis, with each iterated game between a human co-player and the computer as our statistical unit (thus we have 16 independent observations for each of the 2 strong treatments, and 14 independent observations for each of the 2 weak treatments). In the above graph, three stars indicate significance at the level α=0.001, and one star means significance for α=0.05 (using Wilcoxon matched-pairs signed-rank tests with n_ES=n_GS=16, n_EM=n_GM=14). As an auxiliary information, we also provide error bars indicating the 95% confidence interval. Individual results for all 60 individuals are presented in the Supplementary Table 1.