A study of 854 prompts found that four AI models generally produced simpler, less formal emails and cover letters when prompts used language patterns associated with women; the effect was weaker for resignation letters. Gendered names in prompts had virtually no impact.
The models tested were GPT-4, Gemma, Mistral and Llama. Patterns more commonly associated with women included hedges such as “I think,” expressive adjectives, collective terms like “we,” and tag questions. Researchers said mimicking the prompt’s style explained some, but not most, of the differences. The results are due to be presented at the Conference on Language Modeling in October 2026.
The researchers warned that less sophisticated drafts could influence how recipients perceive women’s professional competence and reinforce stereotypes. Lead author Katherine Van Koevering said companies should fix the models rather than put the burden on women to change their speech.
