Evaluating Risk Preferences in GPT Models

Authors: Allen Domingo, Joyce Chen, William Hayes

Field of Study: Science, Technology, Engineering, and/or Math

Program Affiliation: BUPNUR

Faculty Mentors: William Hayes

Easel: 89

Timeslot: Midday

Abstract: The increasing adoption of large language models is significantly impacting decision-making processes, shaping how individuals make informed choices. This study aimed to examine how Generative Pretrained Transformer (GPT) models make decisions under risk compared to the preferences typically observed in humans. This study ran OpenAI’s GPT-3.5 and 4o models through 380 binary choice problems involving a sure option (guaranteed gain or loss) and a gamble with a certain probability of a larger gain or loss. GPT models consistently demonstrated greater risk-seeking behavior in choice problems involving losses, particularly in scenarios where the probability of risk was higher. Differences in model architecture and prior training data may account for the model’s behavior trends. Future research is needed to explore these risk behaviors in a wider range of models to better understand the underlying mechanisms driving decision-making behaviors in large language models.