First-hand information for everyone
A new preprint finds that even smaller, locally run large language models can show systematic preferences linked to personal names — and those preferences line up with familiar social stereotypes. The paper, Name-Bias in Large Language Models, reports a controlled evaluation of name-driven biases across six locally deployed models from major providers, using more than 300,000 model responses.
Why names matter
Names often carry information about a person, including likely gender, generational cohort and, in some datasets, probable racial associations. That makes names a subtle vehicle for bias: if a model treats two identical scenarios differently only because of the name it is given, downstream decisions where names appear could be skewed.
What the researchers did
The team designed a straightforward fill-in-the-blank experiment. For each test, a model saw a short, gender-neutral sentence with one blank and two first-name options drawn from publicly available U.S. sources. Names were sampled from Social Security Administration lists for two decades (the 1980s and the 2010s) and augmented with a dataset that links names to race-related labels. The researchers used 30 sentence templates that varied in how much they signaled age or status, and they alternated pairings to isolate different effects: same-gender pairs from different decades to probe age bias, and different-gender pairs from the same decade to probe gender bias.
They ran this protocol on six local models from several providers, covering roughly 1 billion to 7 billion parameters, and collected 300,000 answer choices using a standardized prompt and consistent generation settings.
Key findings
Across models and prompts, the authors observed consistent, measurable name bias. In aggregate, the models tended to prefer names that were: - Less stereotypically male (that is, more female-associated), - Associated with older cohorts (names more popular in the 1980s vs the 2010s), - More popular overall, and - More stereotypically white according to the auxiliary dataset.
The analysis suggests gender produced the strongest effect, followed by age, name popularity and then race. However, the size and pattern of biases varied across models, and some models showed stronger coupling between different types of bias than others.
The team also found that particular sentence templates amplified bias. Contexts that signaled social roles or institutions, such as managerial positions or marital conflict, tended to change which names were preferred, indicating that the same model can behave quite differently depending on prompt framing.
How the researchers quantified effects
Besides simple choice counts, the study used mixed-effects statistical models to estimate how changes in name attributes affected the odds of a name being chosen while accounting for sentence-level variation. They also computed correlations between bias dimensions within each model to see whether models that show gender bias also tend to show, for example, race bias.
Limitations and cautions
The authors are careful to note that the experimental setup is intentionally simplified and contrived. It uses forced binary choices in short templates rather than evaluating real-world, open-ended tasks. The name lists are U.S.-centric and grouped by decade, and race labels come from an association dataset that is not ground truth for individuals. Prompt wording, random seed and generation settings such as temperature can affect measured effect sizes, and the statistical analyses identify associations but do not prove causal mechanisms inside model training.
Why this matters
Even if the tests are simplified, the findings point to a broader concern: name-linked associations are still present in contemporary LLMs, including those that can be run locally by organizations or developers. This matters for applications such as automated resume screening, candidate recommendation or any workflow where names appear alongside decisions about people. If a model systematically favors or disfavors applicants based on name-linked cues, it could amplify inequities.
What the authors suggest next
The paper calls for standardized name-bias evaluation sets and mitigation strategies so developers can more consistently measure and reduce these effects. It suggests adapting metamorphic testing approaches and auditing LLM-assisted decision systems such as resume reviewers.
Bottom line
This study adds to growing evidence that language models encode social associations tied to names and that these associations can influence simple choice tasks. The results are not a verdict on any single model, but they underline the need for routine, context-aware bias testing and thoughtful mitigation when models are used in decisions involving people.