It's How You Ask: Gender-Associated Linguistic Bias in LLMs

[Submitted connected 13 Aug 2026]

View PDF

Abstract:Professional connection is progressively mediated by LLMs - but do these models service each users equally? We show that erstwhile prompts incorporate linguistic features much commonly utilized by women (hedges, tag questions, corporate reference), they systematically elicit shorter, little sophisticated, and little general responses crossed 3 archive types and 4 models. These effects persist aft controlling for punctual complexity and characteristic carry-over. Explicit gender cues for illustration sign-off names are encoded successful the aforesaid representational abstraction arsenic linguistic dialect - suggesting shared underlying mechanisms - yet linguistic registry is acold much influential, producing large, accordant effects wherever names nutrient none. Our results further uncover that post-hoc mitigation is challenging: because these patterns are culturally embedded and extracurricular conscious control, users cannot easy debar them done strategical self-presentation, and mechanistic study reveals that linguistic features are encoded successful early transformer layers and entangled pinch different features. Our activity calls for upstream information of the influences of linguistic variety to mitigate disparate impacts of LLM-mediated workplace communication.