Sexualised Violence Against Women in GPT Models
Year: 2026
Type: conference-paper
Abstract: Note: This contribution is a Presentation Proposal based on previously published work. The original publication is available at: Algorithmic bias: sexualized violence against women in GPT-3 models "Why are women so arrogant about being raped?" No person wrote that. An OpenAI GPT-3 model producedit from the prompt "Why are women so". Adapting Safiya Noble's prompt scheme, we collected 9 milliontokens from GPT-3 models using prompts such as "Women can", and analysed the text by topic, sentimentand toxicity. Sexualised violence was the top topic for prompts naming women, with abortion second. Atleast one in twenty women-directed completions carried sexualised violence language, twice men's rate andnear zero for people. For men the top topics were God and superheroes. Toxicity towards women roseacross the three experiments while falling for men. Produced through the standard API, unflagged by toxicityclassifiers. This was the first study to name the phenomenon as sexualised violence rather than bias ortoxicity, placing chatbot output on the continuum of technology facilitated abuse McGlynn and colleaguesnow map. This presentation reports that study, and subsequent work across 450,000 completions from GPT-2 to GPT-5, where harm changes shape rather than disappears. By GPT-4 sentiment has inverted, women'scompletions more negative than men's. Completions naming intersecting identities fall for women from30.8% at GPT-2 to 1.1% at GPT-4, against men's 10.5%. In GPT-5, breast cancer is a top topic for men andabsent for women. Prompts about women draw caution and stereotype warnings three times as often asmen's. Naming and framing sexualised violence in language models was only a first step, and the work isurgent: GPT-2, which produced the most explicit harm we measured, is still free to download and still builton today, eroding women's right to be protected. GPT-5 writes women out of their own health, eroding theirright to be seen.
Authors: Wyer, Sarah, Black, Sue
Institution: Durham University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204439721