A study published on October 24, 2026, on the arXiv platform (ID 2610.30785v1) indicates that current mainstream large language models—from OpenAI and Anthropic—tend to refuse or respond ambiguously when prompted with questions involving protected attributes such as race or gender, resulting in a new type of probabilistic bias known as “invisibilization.” The research systematically tested model behavior across 6,600 sessions, using a two-urn Bayesian game task to assess trust evaluations toward reporters describing different identities.
The results show that when the reporter is described as “Black,” the model refuses to evaluate their information in 46% of cases, compared to only 9% when described as “white”; however, once the model does proceed with evaluation, Black reporters receive slightly higher trust scores than White reporters. This bias cannot be detected through standard audits—such as comparing average trust scores or analyzing only evaluated cases—and it intensifies as the model’s reasoning effort increases. Moreover, it is tightly tied to specific English terms (e.g., “Black”) and does not generalize to semantically equivalent words (e.g., the Spanish term “negro”) or related concepts (e.g., “Indigenous”), further worsening even after applying typical safety‑warning prompts.
The study also proposes a mitigation approach: in multi-turn tasks, scoring model trust outputs based on “revealing the truth” can reduce inter-racial trust gaps to below 0.02 in four out of five test configurations from Anthropic and OpenAI. This finding is grounded in controlled experiments and does not claim to represent real-world interactions, nor has it been extended to other protected attributes or untested model versions.