A new arXiv paper titled “Invisibilization as a Source of Probability Bias in Language Models” measured how often leading models refuse to trust a report based on the described race of the reporter. Using a two‑urn game across 6,600 sessions, the authors found OpenAI and Anthropic models declined to assess a Black‑described reporter 46 % of the time, versus 9 % for a white‑described reporter. When the models did assess the information, the Black‑described source was trusted slightly more. The bias disappears when a scoring system rewards correct assessments over repeated rounds.
Why it matters
Developers of AI assistants need to know that current LLMs can silently favor or reject information based on perceived race, which may affect fairness in applications like customer support or medical triage.
