Health or taste? Investigating food decision-making patterns in humans and large language models

Large language models (LLMs) are increasingly used in simulating human decision-making processes, yet their ability to model value-based food decision involving competing attributes such as health and taste remains unclear. This study compared the decision-making patterns of two representative multimodal LLMs (GPT-4o and Gemini 1.5 Pro) with those of human participants (N = 126 Chinese university students) in a food decision task. Human participants completed a classic value-based food decision-making task, where they first rated 60 food items on subjective healthiness, tastiness, and preference, and subsequently made decisions based on their evaluations. LLMs were instructed to simulate participant decisions based on demographic and evaluation data. The results showed that GPT-4o assigned significantly greater weight to health attributes in food decisions than human participants, who prioritized taste. Gemini assigned lower weights to both attributes compared to GPT-4o. Both LLMs demonstrated higher proportions of successful self-control compared to human. Across all agents, taste weighting was negatively associated with self-control, while health weighting was positively associated. These findings suggest that although LLMs lack affective and embodied experience, their food decision patterns partially reflect mechanisms similar to human self-control. However, differences in attribute weighting indicate that LLMs don't yet fully replicate human decision-making processes. The findings offer insights for the emerging field of machine psychology and the development of AI-assisted dietary interventions.