A study published on arXiv on 2026-08-13 shows that prompts containing linguistic features more commonly used by women (hedges, tag questions, collective reference) systematically elicit shorter, less sophisticated, and less formal responses across three document types and four LLMs. Effects persist after controlling for prompt complexity and feature carry-over. Explicit gender cues like sign-off names are encoded in the same representational space as linguistic dialect, but linguistic register produces large, consistent effects while names produce none. Post-hoc mitigation is challenging because patterns are culturally embedded and outside conscious control. Linguistic features are encoded in early transformer layers and entangled with other features.
A research paper on arXiv demonstrates that gender-associated linguistic patterns in prompts cause systematic differences in LLM outputs. Prompts with features more common in women's language lead to shorter, less sophisticated, and less formal responses across multiple document types and models. The effect is driven by linguistic register rather than explicit gender cues, and is difficult to mitigate because it is culturally embedded and encoded early in transformer layers.
The study indicates that linguistic features associated with gender are encoded in early transformer layers and entangled with other features, making post-hoc mitigation difficult. Explicit gender cues like sign-off names share representational space with linguistic dialect, but linguistic register has a much larger effect on output quality. This suggests that bias is not simply a surface-level prompt artifact but is deeply embedded in model representations.
LLM-mediated workplace communication may systematically disadvantage users whose natural language includes hedges, tag questions, or collective references. Since these patterns are outside conscious control, users cannot easily avoid them through strategic self-presentation. This raises concerns about disparate impacts in professional settings where LLMs are used for drafting or editing documents.
Organizations deploying LLMs for communication tasks should be aware that output quality may vary based on the linguistic style of the prompt, potentially affecting employee productivity and fairness. Addressing this bias could improve consistency and equity in AI-assisted writing tools.
Future work may focus on upstream mitigation strategies, such as training data curation or architectural changes to reduce sensitivity to linguistic register. Observable next signals include follow-up studies testing mitigation techniques, model updates addressing such biases, or industry guidelines for inclusive prompt design.