Study finds large language models display stable risk attitudes across tasks

Researchers from multiple institutions introduced a cross-domain framework to quantify how large language models (LLMs) translate perceived risk into actionable decisions. The study tested six representative LLMs alongside 100 human participants across three distinct task domains: spatial navigation, clinical triage, and financial allocation.

Using regression models, the team extracted each agent’s belief-to-decision mapping to measure risk sensitivity and bias. The analysis revealed that most tested LLMs exhibited robust intra-task consistency, meaning their risk decisions remained stable within a fixed task domain. Additionally, the models demonstrated cross-domain rank-order stability, preserving their relative risk posture across different types of tasks.

The findings further indicate that LLMs converge toward a restricted risk-attitude distribution, which differs from the broader variability observed in human baselines. This convergence suggests that LLM risk attitudes may be an intrinsic behavioral property, rather than a task-specific artifact.

The authors argue that understanding these stable risk attitudes is critical for evaluating and aligning AI systems deployed in open-ended, high-stakes environments. They propose that this framework provides a foundation for future research into the origins of such intrinsic behavioral dispositions in LLMs.