GPT Agents in Game Theory Experiments

In this paper, Fulin Guo explores the potential of using Generative Pre-trained Transformer (GPT)-based agents as participants in strategic game experiments. Specifically, Guo focuses on the finitely repeated ultimatum and prisoner's dilemma games, two well-studied games in economics. The author develops 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. For instance, there is a positive relationship between acceptance rates and offered amounts in the ultimatum game and positive cooperation rates in the prisoner's dilemma game. However, some differences between the behavior of GPT and humans are observed in aspects like the evolution of choices over rounds. Moreover, Guo studies 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 for social science research. Overall, this study highlights how AI-based agents such as GPT can be used as experimental subjects for studying human behavior under different conditions. It also suggests that further research could explore more complex strategic interactions involving multiple players or more intricate decision-making processes.

Exploring the Potential of Using Generative Pre-trained Transformer (GPT)-based Agents in Strategic Game Experiments

In this research paper, Fulin Guo explores the potential of using Generative Pre-trained Transformer (GPT)-based agents as participants in strategic game experiments. Specifically, Guo focuses on the finitely repeated ultimatum and prisoner's dilemma games, two well-studied games in economics. The author develops prompts to enable GPT agents to understand the game rules and play the games.

Understanding Game Rules with Prompts

To help GPT agents understand the game rules, Guo developed prompts that explain how each game works and what actions are available to players at each round. For example, for the ultimatum game, a prompt was designed to explain that one player proposes an amount of money to split between them while another player decides whether or not to accept it. Similarly, for the prisoner's dilemma game a prompt was designed to explain that both players can either cooperate or defect with each other at each round.

Results Indicate Realistic Outcomes

The results indicate that given well-crafted prompts GPT can generate realistic outcomes and exhibit behavior consistent with human behavior in certain important aspects. For instance, there is a positive relationship between acceptance rates and offered amounts in the ultimatum game and positive cooperation rates in the prisoner's dilemma game - similar trends were observed when humans played these games as well. However, some differences between GPT’s behavior and humans' were also observed - for example evolution of choices over rounds differed slightly between them.

Treatment Effects Evident In Both Games

Moreover, Guo studied two treatments where GPT agents were prompted either with social preferences or without them before playing both games. The treatment effects were evident in both games - for instance when prompted with social preferences GPT exhibited higher acceptance rates than when prompted without such preferences during ultimatum gameplay; similarly higher cooperation rates were seen during prisoner’s dilemma gameplay when social preferences were included as part of prompting instructions compared to cases where they weren't included . This preliminary exploration indicates that GPT agents can exhibit realistic performance in simple strategic games even if their understanding is limited by prompting instructions alone.

Potential Tool For Social Science Research

Overall this study highlights how AI-based agents such as GPT can be used as experimental subjects for studying human behavior under different conditions which could provide valuable insights into decision making processes across various scenarios including those involving multiple players or more intricate decision making processes than those explored here by Guo.. It also suggests that further research could explore more complex strategic interactions involving multiple players or more intricate decision making processes which could potentially lead us closer towards understanding why people behave differently under different circumstances better than ever before!