Coupled social dilemmas reveal positive cross-game transfer in humans and contrasting cue use in tested GPT agents

Cooperation across social settings requires deciding whether to reciprocate a partner’s helpful behavior or exploit it when incentives change. We examined this tension in an online experiment with 430 participants playing a Prisoner’s Dilemma (PD), either alone or followed by a Snowdrift game (SD), where defection pays more against a sufficiently likely cooperator. We also placed programmed bots—computer-controlled players following a fixed rule—in some groups. By programming them to cooperate in PD and reciprocate their partner’s PD action in SD, we created a consistent source of cooperation and reciprocity. This allowed us to ask whether humans also cooperated more with one another. Humans cooperated more in SD after their partner cooperated in PD, even when observed partner behavior made defection the better immediate material response. This pattern is consistent with reciprocity, although choices do not identify beliefs or motives. Human–human cooperation was higher in groups containing bots; coupling alone produced no detected increase in mean PD cooperation. Separate experiments with GPT language-model agents asked whether frequent cooperation was accompanied by similar responses across games. Despite high PD cooperation, partner PD cooperation predicted less GPT cooperation in SD in the primary pooled analysis. This association varied across configurations and did not generally indicate payoff-maximizing choices. Simulations using fitted human response rules illustrated the possible effects of alternative bot policies. These findings reveal responses to cooperative partners across contexts that overall cooperation rates can obscure.