A new preprint on arXiv highlights names as a subtle but strong target for bias in large language models. Because names carry demographic indicators such as age, gender, and race, they can influence model outputs in ways that are easy to overlook.

The paper presents an evaluation of name-based bias across six local large language models. The abstract describes the scope of the study, but does not include results or conclusions, so the actual extent of bias in these models remains to be seen.

This work points to the need for careful auditing of LLMs beyond obvious demographic categories, since even a simple name can trigger biased behavior.