Refusal as silence: Gendered disparities in Vision-Language Model responses

Abstract

Refusal behavior by Large Language Models (LLMs) is increasingly visible in content moderation, yet little is known about how refusals vary by the identity of the user making the request. This study investigates refusal as a sociotechnical outcome through a counterfactual persona design. Focusing on a Vision-Language Model (GPT-4V), we examine how gendered persona in prompts influence refusal in binary gender classification tasks. We vary gender identity across male, female, non-binary, and transgender personas while keeping the classification task and visual input constant. We find that transgender and non-binary personas experience significantly higher refusal rates, even in non-harmful contexts. Our findings also provide methodological implications for equity audits using LLMs. We underscore the importance of modeling identity-driven disparities and caution against uncritical use of artificial intelligence systems for content coding. This study advances algorithmic fairness by reframing refusal as a communicative act that may unevenly regulate epistemic access and participation.