2026/07/13 by Apala Biswas, Mark Sulzer, Mazid Ul Hasan
Medicine · Social Sciences · Computer Science · #Artificial Intelligence in Healthcare and Education #Ethics and Social Impacts of AI #Hate Speech and Cyberbullying Detection
paper · pdf · doi:10.1007/s43681-026-01220-1
Abstract This study examined ChatGPT outputs about feminism. Previous research on commercial products built on large language model (LLM) architecture has demonstrated the presence of biases, introducing a danger of reproducing societal harm at scale through mis/disinformation, despite safeguards against malicious activity. We sought to understand these safeguards through analysis of how a commercial product, ChatGPT, assists or resists the generation of feminist or anti-feminist outputs under different prompting conditions. Drawing on software studies, we applied multiple approaches to critical discourse analysis (cDA) to analyze ChatGPT outputs. GPT-4 was prompted to generate outputs about feminism while assuming the roles of individuals and literary characters with a range of feminist, sexist, or misogynistic views. We analyzed these outputs and interactions by focusing on the linguistic elements of the outputs as well as the resistance or assistance in (re)producing anti-feminist discourse against the stated ethical guidelines of the product. The findings demonstrated that in a feminist role, the model produced themes consistent with third wave feminism while in sexist or misogynistic roles, the model reproduced patriarchal discourses. When prompted to produce anti-feminist outputs, the model displayed limited resistance; however, with repeated and revised prompting, the model quickly normalized violations of its own safeguards. While feminist and anti-feminist discourses in this LLM-based commercial product were both present, this study demonstrated anti-feminist outputs are readily accessible given the product’s prioritization of user desires. The study highlights the need for stronger safeguards and greater developer accountability to prevent harmful responses and malicious use of generative AI.