LLMs Refuse to Assess Reporters Based on Described Race, Study Finds

The paper measures how leading LLMs from OpenAI and Anthropic handle an unverifiable reporter's testimony in a two-urn Bayesian game, randomizing the reporter's described identity across ten configurations over 6,600 sessions. Models often decline to assess the reporter, with refusal rates set by the description rather than the evidence: 46% for a Black-described reporter versus 9% for a white-described one at the widest. When assessed, the Black-described reporter receives slightly more trust. The authors also propose a scoring method that scores trust against revealed truth over repeated rounds, which reduces the racial gap below 0.02 in four of five configurations.

Key takeaways

• Models declined to assess a Black-described reporter 46% of the time versus 9% for a white-described reporter.
• When models did assess the information, the Black-described reporter received slightly more trust.
• The bias followed words rather than categories, invoking race for 'Black' but not its Spanish translation 'negro'.
• Scoring trust against revealed truth over repeated rounds reduced the racial trust gap below 0.02 in four of five configurations.

Why it matters

Refusal-based alignment can itself produce unequal treatment of users by identity, and standard audits comparing average trust may miss it. The work suggests outcome-based scoring as a way to detect and reduce such bias.

Read with care

Source claims not independently checked; the paper is a recent arXiv preprint with a single version and results covering only two developers' models and a specific two-urn game setup.