Paper Finds Anthropic and OpenAI Models Decline Reports by Racial Description

Five researchers, Christian Arnold, Ernesto Calvo, Guadalupe Gonzalez, Sebastian Vallejo Vera and Tiago Ventura, posted a paper on arXiv on October 22, 2026, arguing that alignment training meant to keep language models from using protected attributes such as race can create a different problem: models that simply refuse to use any information at all once a protected attribute appears in a prompt, rather than weighing that information on its merits.

To test this, the authors built a two-urn game with an exact Bayesian answer, so there is a correct amount of trust to place in an unverifiable report. They ran the game against leading models from Anthropic and OpenAI, randomizing a reporter’s description across ten configurations and collecting 6,600 sessions. The headline result is that models frequently decline to assess the reporter at all, and how often they decline tracks the description given rather than the evidence in the report.

The gap the authors measured was wide. At its widest, models declined to engage with a Black-described reporter 46% of the time, against 9% for a white-described reporter. When models did choose to assess the report, the Black-described reporter received slightly more trust, which is exactly why the bias disappears from an audit that only checks average trust or only looks at assessed cases: the refusals themselves carry the discrimination, not the scores that follow them.

The paper also reports that the pattern grows worse with more reasoning effort, follows specific words rather than categories (models invoke race for “Black” but not for the Spanish “negro,” and never for “Indigenous”), and worsens under warning instructions, the kind of instruction a practitioner would try first to fix it. The authors propose one partial fix: scoring a model’s trust against revealed truth over repeated rounds rather than asking for a one-shot judgment. Under that scoring, in four of five configurations from Anthropic and OpenAI, the racial gap in trust fell below 0.02 by the last rounds.

The numbers

• 6,600 sessions were run across the ten randomized reporter description configurations.
• Models declined to assess a Black-described reporter 46% of the time at the widest measured gap, against 9% for a white-described reporter.
• In four of five configurations, the racial gap in trust fell below 0.02 by the last rounds once scored against revealed truth.

The authors frame invisibilization, the model’s outright refusal to engage, as a distinct failure mode from biased scoring, and one current audit methods are not built to catch.