Implicit bias in safety-aligned large language models: A multi-faceted evaluation of clinical decision-making and health equity
Qiufeng Jia, Yuhang Wen, Yuyan Liu, Hui Zhao, Qiongge Yu, Yu Long, Dan Sun, Yufeng Yu · 2026
Read Implicit bias in safety-aligned large language models: A multi-faceted evaluation of clinical decision-making and health equity by Qiufeng Jia, Yuhang Wen, Yuyan Liu, Hui Zhao, Qiongge Yu, Yu Long, Dan Sun, Yufeng Yu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/plos-pone-0348819.txt.
Description
Qiufeng Jia and colleagues probed ten leading AI language models for hidden bias in medical decisions. Models can pass open bias tests while still carrying quiet associations between social groups and traits. The team built twenty-four medical test sets spanning gender, race, wealth, health status, religion, and health systems, then measured bias three ways, including word association and comparative decision tasks.
The evaluation covers both commercial systems and open models from several countries. Implicit bias in a tool that drafts advice or triages patients could steer real care unfairly, even when no explicit prejudice appears. A multi-method audit like this shows where the associations live and how strongly they move decisions, giving developers and hospitals concrete targets for correction.
Catalog record
| CREATOR | Qiufeng Jia, Yuhang Wen, Yuyan Liu, Hui Zhao, Qiongge Yu, Yu Long, Dan Sun, Yufeng Yu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49780 |
| LICENSE | CC_BY |
| RIGHTS | CC BY |
| SOURCE | PLOS |