LLMs Give Shorter, Less Sophisticated Answers to Prompts Written in Women's Linguistic Register — and Explicit Gender Cues Change Nothing
Submitted 2026-08-13 (arXiv 2608.13328), this study across four LLMs finds that prompts containing hedges, tag questions, and collective references — linguistic features more commonly used by women — systematically elicit shorter, less sophisticated, and less formal responses, controlling for prompt complexity and feature carry-over. The sharp result is the null: explicit gender cues such as sign-off names produced no measurable effect, while linguistic register produced large, consistent effects. Mechanistic analysis places the encoding early in the transformer layers, entangled with other features, which is why the authors argue it resists mitigation. The practical implication for anyone building on LLMs is that output-quality disparities can be invisible to any fairness test that keys on demographic markers rather than phrasing.
↳ Follow the thread