TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

safety

A Biased Turn Changed the Next Answer

Six of eight tested models expressed more bias after exposure to biased conversational reasoning.

Published Updated Story ID: mp-2026-08-09-019
Read the complete editionStory JSON

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

  1. 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

  1. arXiv preprint 2608.05166arXiv · primary research

Corrections

No corrections have been recorded for this story.