GPT Agents in Game Theory Experiments

Résumé : This paper explores the potential of using Generative Pre-trained Transformer (GPT)-based agents as participants in strategic game experiments. Specifically, I focus on the finitely repeated ultimatum and prisoner's dilemma games, two well-studied games in economics. I develop prompts to enable GPT agents to understand the game rules and play the games. The results indicate that, given well-crafted prompts, GPT can generate realistic outcomes and exhibit behavior consistent with human behavior in certain important aspects, such as positive relationship between acceptance rates and offered amounts in the ultimatum game and positive cooperation rates in the prisoner's dilemma game. Some differences between the behavior of GPT and humans are observed in aspects like the evolution of choices over rounds. I also study two treatments in which the GPT agents are prompted to either have social preferences or not. The treatment effects are evident in both games. This preliminary exploration indicates that GPT agents can exhibit realistic performance in simple strategic games and shows the potential of using GPT as a valuable tool in social science research.