Some Large Language Models Exhibit Consistent Risk Attitudes
06:00 · July 21, 2026 · arXiv cs.AI RSS
As artificial intelligence systems are deployed in open-ended, high-stakes settings, a critical dimension remains unmeasured: how perceived risk is translated into action. We test whether large language models (LLMs) exhibit systematic and consistent risk attitudes under uncertainty. We introduce a cross-domain framework that decouples contextual risk belief from categorical decision, and apply it to six representative LLMs and 100 human participants across spatial navigation, clinical triage, and financial allocation tasks. Using regression models, we extract each agents belief-to-decision mapping and quantify risk sensitivity and risk attitude bias. We find that most tested LLMs exhibit (i) robust intra-task consistency, indicating stable mappings from contextual belief to risk decision within a fixed task domain; (ii) cross-domain rank-order stability, preserving relative risk posture across tasks; and (iii) a convergence toward a restricted risk-attitude distribution relative to the broader human baseline. These results reveal risk attitude as a stable and previously uncharacterized dimension of LLM behavior, establishing a foundation for evaluating and aligning AI systems in open-ended decision-making and motivating further investigation into the origins of these intrinsic behavioral dispositions.
Summary
This primary research paper examines whether large language models develop stable risk attitudes when translating perceived uncertainty into decisions. The authors introduce a cross-domain framework that isolates the mapping from contextual risk belief to categorical action, separating this process from factual perception or probability estimation. They apply the approach to six representative LLMs and a baseline group of 100 human participants, using ordered logistic regression to derive two quantitative indices: risk sensitivity, which measures responsiveness to rising perceived risk, and risk attitude bias, which captures systematic deviation from neutral decision patterns.
Three structurally distinct tasks were used to test consistency: drone navigation under uncertain wind and obstacles, clinical triage based on evolving physiological signals, and financial portfolio allocation under stochastic market conditions. Each task required agents to form an intermediate contextual belief from sequential observations before selecting a risk-related action. The design enabled direct comparison of intra-task stability, where models map similar beliefs to similar decisions within one domain, and cross-domain rank-order stability, where relative risk posture is preserved across dissimilar settings.
Results indicate that most tested models display robust intra-task consistency and maintain their relative risk posture across all three domains. At the same time, the LLMs converge on a markedly narrower distribution of risk attitudes than the human sample, suggesting that current training regimes produce more uniform behavioral dispositions than those observed in people. The authors conclude that risk attitude constitutes an intrinsic, reproducible dimension of LLM behavior that current capability benchmarks overlook, with direct implications for alignment and safety evaluation in high-stakes, open-ended decision environments.
Why it matters
This research is highly relevant for Dutch AI researchers and policymakers focused on ethical and transparent AI, as it provides a novel framework for auditing the intrinsic risk behaviors of LLMs. Understanding these latent risk profiles is crucial for deploying AI in high-stakes environments and aligns perfectly with the EU's stringent risk management requirements.
