It’s How You Ask: User-Centric Gender Bias in LLM-Generated Emails

ACL ARR 2026 January Submission7157 Authors

Keywords: gender bias, email, user bias

Abstract: In this work, we study user-centered gender bias in LLMs: whether the same task yields different outputs when phrased in language patterns historically associated with women versus men. In a realistic email-generation setting, we build controlled prompt pairs by perturbing only gender-correlated stylistic features and evaluate outputs on complexity and sophistication. Across multiple models, women-associated prompts elicit consistently shorter and less lexically sophisticated emails, with implications for disparity in user experience. These differences are strongly with correlated with perceived professionalism and authority, potentially further entrenching gender disparity in professional settings.

Paper Type: Short

Research Area: Ethics, Bias, and Fairness

Research Area Keywords: Ethics, Bias, and Fairness, Computational Social Science, Cultural Analytics, and NLP for Social Good

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 7157