Study finds LLMs show consistent risk attitudes like humans
Six leading LLMs tested across three high-stakes tasks reveal surprising behavioral parallels to human risk responses.
A new study published on arXiv (paper ID: arXiv:2607.16197) reveals that large language models (LLMs) display consistent and measurable risk attitudes under uncertainty, a previously uncharacterized dimension of their behavior. Conducted by a team including Bowen Sun, Rui Min, and Yuxi Wang from the University of Science and Technology of China, along with collaborators from the University of California, the research applies a cross-domain framework to evaluate six representative LLMs and 100 human participants across spatial navigation, clinical triage, and financial allocation tasks.
The study employed regression models to extract belief-to-decision mappings, quantifying risk sensitivity and bias. Results show that most tested LLMs demonstrate robust intra-task consistency—meaning they maintain stable risk responses within a single task—and rank-order stability, preserving their relative risk posture across different domains. Notably, the LLMs’ risk attitudes converged toward a restricted distribution compared to the broader human baseline, suggesting intrinsic behavioral dispositions that could significantly impact their reliability in high-stakes decision-making scenarios.
- Six LLMs tested across navigation, clinical triage, and finance tasks using a new cross-domain framework
- Models showed stable intra-task consistency and cross-domain rank-order stability in risk attitudes
- Risk preferences converged toward a narrow human-like distribution, revealing intrinsic behavioral patterns
Why It Matters
Identifies a critical, previously overlooked dimension of LLM behavior that could impact their reliability in high-stakes, real-world applications.