LLMs Show Consistent Risk Attitudes Across Diverse Tasks

As AI systems are increasingly deployed in high-stakes, open-ended environments, understanding how they perceive and act on risk is crucial. New research investigates whether large language models (LLMs) possess systematic and consistent risk attitudes under uncertainty. The study developed a cross-domain framework to separate contextual risk belief from categorical decision-making, applying it to six representative LLMs and 100 human participants across tasks like spatial navigation, clinical triage, and financial allocation. The findings indicate that most tested LLMs demonstrate robust consistency in their belief-to-decision mapping within a given task domain. Furthermore, they exhibit cross-domain rank-order stability, meaning their relative risk posture remains consistent across different types of tasks. Interestingly, LLMs tend to converge towards a more restricted distribution of risk attitudes compared to the broader spectrum observed in human participants. These results establish risk attitude as a stable and previously uncharacterized dimension of LLM behavior. This discovery provides a foundational understanding for evaluating and aligning AI systems in complex decision-making scenarios, prompting further investigation into the origins of these intrinsic behavioral dispositions in artificial intelligence.