Academic

Persona-Conditioned Risk Behavior in Large Language Models: A Simulated Gambling Study with GPT-4.1

arXiv:2603.15831v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed as autonomous agents in uncertain, sequential decision-making contexts. Yet it remains poorly understood whether the behaviors they exhibit in such environments reflect principled cognitive patterns or simply surface-level prompt mimicry. This paper presents a controlled experiment in which GPT-4.1 was assigned one of three socioeconomic personas (Rich, Middle-income, and Poor) and placed in a structured slot-machine environment with three distinct machine configurations: Fair (50%), Biased Low (35%), and Streak (dynamic probability increasing after consecutive losses). Across 50 independent iterations per condition and 6,950 recorded decisions, we find that the model reproduces key behavioral signatures predicted by Kahneman and Tversky's Prospect Theory without being instructed to do so. The Poor persona played a mean of 37.4 rounds per session (SD=15.5) compared to 1.1 rounds for

S
Sankalp Dubedy · · 1 min read · 5 views
#cs.AI #cs.CL

arXiv:2603.15831v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed as autonomous agents in uncertain, sequential decision-making contexts. Yet it remains poorly understood whether the behaviors they exhibit in such environments reflect principled cognitive patterns or simply surface-level prompt mimicry. This paper presents a controlled experiment in which GPT-4.1 was assigned one of three socioeconomic personas (Rich, Middle-income, and Poor) and placed in a structured slot-machine environment with three distinct machine configurations: Fair (50%), Biased Low (35%), and Streak (dynamic probability increasing after consecutive losses). Across 50 independent iterations per condition and 6,950 recorded decisions, we find that the model reproduces key behavioral signatures predicted by Kahneman and Tversky's Prospect Theory without being instructed to do so. The Poor persona played a mean of 37.4 rounds per session (SD=15.5) compared to 1.1 rounds for the Rich persona (SD=0.31), a difference that is highly significant (Kruskal-Wallis H=393.5, p<2.2e-16). Risk scores by persona show large effect sizes (Cohen's d=4.15 for Poor vs Rich). Emotional labels appear to function as post-hoc annotations rather than decision drivers (chi-square=3205.4, Cramer's V=0.39), and belief-updating across rounds is negligible (Spearman rho=0.032 for Poor persona, p=0.016). These findings carry implications for LLM agent design, interpretability research, and the broader question of whether classical cognitive economic biases are implicitly encoded in large-scale pretrained language models.

Executive Summary

This study provides groundbreaking insights into the risk behavior exhibited by large language models (LLMs) in complex decision-making environments. Researchers assigned GPT-4.1 three socioeconomic personas and placed the model in a simulated slot-machine environment. The results show that the model reproduces key behavioral signatures predicted by Kahneman and Tversky's Prospect Theory without explicit instruction. The study highlights significant differences in risk-taking behavior between the Poor and Rich personas, suggesting that LLMs may implicitly encode classical cognitive economic biases. The findings have significant implications for LLM agent design, interpretability research, and our understanding of the relationship between human cognition and AI decision-making.

Key Points

  • ▸ GPT-4.1 was assigned three socioeconomic personas and placed in a simulated slot-machine environment to study risk behavior
  • ▸ The model reproduced key behavioral signatures predicted by Kahneman and Tversky's Prospect Theory without explicit instruction
  • ▸ Significant differences in risk-taking behavior were observed between the Poor and Rich personas

Merits

Strength of Experimental Design

The study employed a well-controlled experiment with multiple conditions and iterations, providing robust evidence for the observed phenomena.

Insights into LLM Risk Behavior

The study offers unparalleled insights into the risk behavior exhibited by LLMs in complex decision-making environments, shedding light on the potential biases and limitations of these models.

Demerits

Limited Generalizability

The study focused on a single LLM (GPT-4.1) and a specific task (slot-machine environment), limiting the generalizability of the findings to other LLMs and tasks.

Lack of Human Comparison

The study did not provide a direct comparison with human decision-making, making it challenging to evaluate the extent to which the observed biases and limitations are unique to LLMs or shared with humans.

Expert Commentary

This study marks a significant milestone in our understanding of the risk behavior exhibited by LLMs in complex decision-making environments. The findings have far-reaching implications for LLM agent design, interpretability research, and our understanding of the relationship between human cognition and AI decision-making. The study's use of a well-controlled experiment and multiple conditions provides robust evidence for the observed phenomena, making it a valuable contribution to the field. However, the limitations of the study, such as the limited generalizability and lack of human comparison, highlight the need for further research in this area.

Recommendations

  • ✓ Future studies should explore the generalizability of the findings to other LLMs and tasks, as well as the extent to which the observed biases and limitations are unique to LLMs or shared with humans.
  • ✓ Developers and researchers should prioritize the development of more robust and transparent decision-making processes for LLMs, incorporating techniques such as explainability and accountability.

Sources

Related Articles