Playing Prisoner's Dilemma Games with a Large Language Model
Andreas Orland; Kazuhiro Takemoto; Philipp Külpmann
Working paper (2026). Version: 1 May 2026.
Abstract
Large Language Models (LLMs) are increasingly participating in strategic environments. We provide a behavioral characterization of ChatGPT-3.5-turbo in a one-shot Prisoner’s Dilemma game. Using a full factorial experimental design, we independently vary payoff parameters, the number of interaction partners, the strategy space, and the elicitation order, while eliciting both decisions and stated beliefs. We document three central patterns. First, the model exhibits high baseline cooperation despite pessimistic expectations about others. Second, cooperation is largely insensitive to payoff variation but systematically responds to structural and procedural manipulations such as group size, action space, and elicitation order. Third, the gap between beliefs and actions is condition-dependent. We interpret these findings as a behavioral profile of an aligned LLM operating in a strategic environment. The results suggest that interaction protocols and institutional structure may shape outcomes more strongly than marginal incentive adjustments when such agents are involved.Links and resources
Keywords: Prisoner’s Dilemma; Cooperation; Experiment; Algorithms; Large Language Models