Students

This project investigates how different generative AI systems behave when engaging in strategic interactions through game-theoretic scenarios, examining whether these tools exhibit emergent cooperative or competitive behaviors that mirror human decision-making patterns. Drawing inspiration from recent research on algorithmic collusion by large language models, we will systematically test how various AI systems “play” classic games such as the prisoner’s dilemma, public goods games, and coordination games under different contextual framings and stake levels. The core research question centers on how different language models play these games, and in particular how they perform on different parts of the game (e.g. coming up with strategies, following strategies, evolving strategies from past information).  

The experimental design will manipulate both the contextual framing of games (abstract versus concrete scenarios like market competition or environmental cooperation) and the perceived stakes involved (high-value versus low-value outcomes) to understand how these factors shape AI decision-making. By comparing outcomes across different generative AI architectures and training paradigms, this project aims to uncover whether certain AI systems are more prone to cooperative versus competitive strategies, and whether these tendencies can be predicted based on their training data or architectural features. The findings will have important implications for understanding potential risks and benefits as AI systems increasingly interact with each other and humans in strategic settings, particularly in economic contexts where coordination or competition could have real-world consequences.