To Prompt a Masculine Woman: Mapping Biases on the Gender Spectrum

AI audits are meant to detect unfair outputs, yet they frequently dismiss queer genders. Drawing on queer and participatory research, this paper proposes a methodology for systematically incorporating binary and nonbinary genders. Focusing on a text-to-image AI model, we generated 44 portraits with Midjourney by prompting the labels of the Gender Diamond interactive visualisation of the gender spectrum. Then, we conducted eight workshops with 60 queer and queer-supportive participants. They individually evaluated the portraits in terms of accuracy — how well the portraits aligned with their gender stereotypes — and discrimination — whether they perceived problematic associations. Next, participants collectively discussed the elicited sociotechnical status quo and speculated on alternative mitigation strategies when prompting for gendered portraits. Our findings are clustered in three areas. First, regarding accuracy, we found that Midjourney conveys a reliable, broad representation of who conforms to the binary and who does not. However, it is inaccurate with specific nonbinary genders. Second, regarding discrimination, we identified cisnormative, racist, ableistic, reproductivist, ageist, and fatphobic assumptions, mapped across the gender spectrum. Third, we propose three queer scenarios for AI-generated gendered portraits focused on contextual appropriateness. By exploring its advantages and limits, we propose the Gender Diamond as a tool to expand and guide critical engagement with power and technology.