Name-Bias in Large Language Models

Summary

Researchers evaluated name-based bias in 6 local large language models from multiple providers and parameter scales. Using a fill-in-the-blank task with balanced first-name pairs from U.S. Social Security Administration data and a race classification dataset by Rosenman et al., they ran 300,000 responses and found consistent bias: models preferred names that are less stereotypically male, older, popular, and white. Gender was the strongest effect, followed by age, name popularity, and race.