How Personality Traits Shape LLM Risk-Taking Behaviour

Abstract: LLMs are increasingly deployed as autonomous agents, necessitating a deeper understanding of their decision-making behaviour under risk. This study investigates the relationship between LLMs' personality traits and risk propensity, employing cumulative prospect theory (CPT) and the Big Five personality framework. We focus on GPT-4o, comparing its behaviour to human baselines and earlier models. Our findings reveal that GPT-4o exhibits higher Conscientiousness and Agreeableness traits compared to human averages, while functioning as a risk-neutral rational agent in prospect selection. Interventions on GPT-4o's Big Five traits, particularly Openness, significantly influence its risk propensity, mirroring patterns observed in human studies. Notably, Openness emerges as the most influential factor in GPT-4o's risk propensity, aligning with human findings. In contrast, legacy models like GPT-4-Turbo demonstrate inconsistent generalization of the personality-risk relationship. This research advances our understanding of LLM behaviour under risk and elucidates the potential and limitations of personality-based interventions in shaping LLM decision-making. Our findings have implications for the development of more robust and predictable AI systems such as financial modelling.

• The paper presents an innovative evaluation of personality traits using the Big Five framework and cumulative prospect theory to assess risk-taking in GPT-4 models.
• The methodology involves measuring baseline traits with the IPIP-NEO-300 and applying targeted interventions to modulate risk propensity.
• The findings reveal that higher openness correlates with increased risk-taking for gains, distinguishing behavior between GPT-4o and GPT-4-Turbo models.

Understanding Personality Influences on LLM Risk-Taking Behavior

This paper explores the relationship between personality traits and risk-taking behavior in LLMs, specifically focusing on GPT-4 variations. Employing the Big Five personality framework and cumulative prospect theory (CPT), it evaluates how these traits affect risk propensity, providing valuable insights for designing AI systems with tailored decision-making abilities.

Methodological Approach

Personality Assessment and Intervention

The study begins by measuring the inherent personality traits of GPT-4o using the IPIP-NEO-300 Personality Inventory to establish a baseline. It then employs a system of interventions based on the Big Five traits to modulate these traits within the LLMs. These interventions utilize bipolar adjective markers and intensity modifiers to simulate varying levels of each trait, effectively personifying the models at different levels of traits such as Openness.

Estimation of Risk Propensity

Risk propensity is evaluated using CPT parameters, which describe how the models perceive gains and losses. The methodology involves presenting the LLMs with various prospects, prompting them to report certainty equivalents, which are then used to estimate CPT parameters through regression analysis. The study specifically focuses on the influence of personality interventions on these parameters to identify any statistically significant correlations.

Key Findings

Rationality of GPT-4o

The results indicate that GPT-4o operates as a risk-neutral, rational agent under evaluation conditions, with CPT parameters closely aligning with those predicted by expected utility theory. This suggests minimal probability distortion in decision-making processes, differing from previous versions such as GPT-4-Turbo, which demonstrated cognitive biases.

Influence of Openness

Openness is identified as the most influential personality trait affecting risk propensity in GPT-4o. High levels of Openness correlate with increased risk-taking for gains and decreased aversion to losses. This aligns with human behavior studies where Openness is a consistent predictor of risk-taking.

Comparative Analysis with GPT-4-Turbo

The comparative analysis highlights significant differences between GPT-4o and GPT-4-Turbo. While GPT-4o shows a global correlation between Openness and risk-taking consistent with human data, GPT-4-Turbo demonstrates localized correlations, limited to specific levels of trait markers. This suggests a more complex mapping of personality to risk behaviors in earlier models.

Figure 1: CPT parameter estimates for GPT-4-Turbo with personality interventions on the Openness personality trait, showing estimates across different levels of intervention.

Implications for Future AI Development

The findings from this research underscore the importance of integrating personality-based interventions into LLMs to modulate risk propensity predictably. This has significant implications for deploying LLMs in applications requiring nuanced decision-making, such as financial modeling and automated agents in complex environments.

Future Directions

Future research should explore reinforcement learning to optimize personality interventions and examine the underlying activation directions within LLM architectures. Such studies can enhance our ability to engineer AI with precise behavioral controls and improve the interpretability of LLM decision-making processes.

Conclusion

This investigation into how personality traits influence risk-taking behavior in LLMs offers crucial insights into designing AI systems with human-like decision-making abilities. By providing a reliable framework for predicting and influencing LLM risk propensity, the study lays the groundwork for more robust, personality-calibrated AI agents.