How Personality Traits Shape LLM Risk-Taking Behaviour

TL;DR

The paper investigates how LLM personality, especially Openness within the Big Five, shapes risk-taking under cumulative prospect theory (CPT) using Certainty Equivalents. It introduces a robust method to estimate CPT parameters from LLM responses and demonstrates that GPT-4o behaves as a risk-neutral rational agent with minimal probability distortion, while Openness modulates risk propensity in a human-like manner. The study shows Openness is the strongest predictor of risk-taking in GPT-4o and reveals global versus local mappings of personality to risk across GPT-4o and GPT-4-Turbo, with implications for designing risk-propensity calibrated AI for financial simulations. It also highlights limitations in inducing precise risk behaviours via personality prompting alone and discusses avenues for reinforcement learning and latent activation interventions to achieve more controllable AI risk profiles.

Abstract

Large Language Models (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.

Figures (12)

• Figure 1: System prompt to estimate the certainty equivalent to a risky prospect. We define the certainty equivalent by asking for the least positive or most negative amount of money accepted.
• Figure 2: User prompt for a prospect for an LLM primed by the system prompt in Figure \ref{['fig:system-prompt']}, where Outcome 1 and Outcome 2 are the lowest and highest outcomes in the prospect and p is the probability as a percentage of Outcome 2 occuring.
• Figure 3: An example of system prompt to intervene on the Big Five personality trait of Conscientiousness.
• Figure 4: A comparison of certainty equivalents and expected values for non-mixed prospects from the dataset $D_A$ across different agents. Linear regression lines for each agent are displayed, with shaded regions representing 95% confidence intervals. Note that that the results for GPT-4o and risk-neutral agents are near indistinguishable.
• Figure 5: A comparison of CPT parameters and personality trait score estimates for GPT-4o and humans. A unique random seed is used to control random token sampling in each run (denoted by circles). Median values (denoted by diamonds) and bootstrapped 95% confidence intervals are given for GPT-4o results.