Artificial Intolerance: Stigmatizing Language in Clinical Documentation Skews Large Language Model Decision-Making

In three linesarXiv study demonstrating that 9 frontier LLMs inherit stigmatizing language bias from clinical notes. Models reduce treatment aggressiveness when exposed to a single stigmatizing sentence. Chain-of-Thought and self-debiasing show limited mitigation efficacy.

Summary generated by Claude — human-verified