LLMs show second-order bias when judging biased text: EMNLP study

An EMNLP 2026 Findings paper from Toronto and Vector Institute finds that when asked who would accept a biased text, LLMs refuse to name an attributed group only 0.4% of the time, revealing systematic second-order bias.

2026-10-27 ~ 2026-10-27 · 2 related posts

• LLM bias judges name groups instead of saying 'unknown', refusing only 0.4% of the time — ZhijingJin · 2026-10-27

1 near-duplicate retellings: ZhijingJin