Moral Hazard in Multi-Agent Language Models

This paper adapts Holmström’s team moral-hazard model into the Dialogue Moral Hazard Game, a controlled textual environment where agents may pay a query cost to reveal hidden safety information that mainly benefits another agent. It evaluates nine open-weight models and one frontier API model across querying, information transfer, local-reward preservation, unsafe choices, format validity, and team success. The abstract reports that base open-weight models often preserve local reward without achieving team success, or query without transmitting decision-changing information. GPT-5.6 Sol reaches ceiling behavior in the primary setting, with a query threshold closely tracking the theoretical private-share boundary in 3,015 scripted-partner decisions.