Facing the grim truth: Repeated prisoner’s dilemma against robot opponents

Cooperation in repeated interactions is important for much socio-economic activity. In this paper we put subjects in the simplest dynamic setting that can rationalize cooperative behavior while eliminating confounding factors such as multiple equilibria, strategic uncertainty, and other regarding concerns. We find that, over all supergames, only 1-2% of subjects behave perfectly consistently with rational choice predictions, and only 3-5% behave consistently with the theory at least 95% percent of the time. These low frequencies essentially amount to noise, and suggest that the rational choice framework used to explain cooperative behavior may not be empirically relevant, either in the standard subject pool or in a more representative online subject pool. We document that while most subjects make dominated choices resulting in money left on the table, a substantial minority are able to achieve higher total points than predicted by rational choice theory.