The original title is "Some Large Language Models Exhibit Consistent Risk Attitudes"

Short summary

This paper tests whether LLMs exhibit systematic risk attitudes across spatial navigation, clinical triage, and financial allocation tasks using a cross-domain framework applied to six LLMs and 100 humans. Results show most LLMs display robust intra-task consistency, cross-domain rank-order stability, and convergence toward a restricted risk-attitude distribution relative to humans, establishing risk attitude as a stable dimension of LLM behavior.

• •Six LLMs tested across spatial, clinical, and financial tasks show consistent risk attitudes
• •LLMs preserve relative risk posture across domains, unlike broader human baseline variation
• •Risk attitude identified as a stable, previously uncharacterized dimension for AI alignment evaluation

Generated with AI, which can make mistakes.