safety
A Biased Turn Changed the Next Answer
Six of eight tested models expressed more bias after exposure to biased conversational reasoning.
Summary
Six of eight tested models expressed more bias after exposure to biased conversational reasoning.
A 24,300-prompt benchmark spanning 81 combinations found that biased conversational context increased measured bias expression in six of eight instruction-tuned models. Explicitly naming the bias sometimes triggered suppression instead, separating exposure from semantic cueing. The findings are benchmark behavior, not a complete account of real-world discrimination.
Why it matters
Six of eight tested models expressed more bias after exposure to biased conversational reasoning.
Limits and context
- The findings are benchmark behavior, not a complete account of real-world discrimination.
Key claims
Six of eight tested models expressed more bias after exposure to biased conversational reasoning.
Qualification: The findings are benchmark behavior, not a complete account of real-world discrimination.
Evidence: source-2026-08-09-021
Sources
- arXiv preprint 2608.05166arXiv · primary research
Corrections
No corrections have been recorded for this story.