"It's Complicated": Negotiating Accessibility and (Mis)Representation in Image Descriptions of Race, Gender, and Disability
Cynthia L. Bennett, Cole Gleason, Morgan Klaus Scheuerman, Jeffrey P. Bigham, Anhong Guo, Alexandra To · 2021 · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3411764.3445498
Summary
This qualitative study investigates how screen reader users who are also Black, Indigenous, People of Color (BIPOC), non-binary, and/or transgender navigate the complex landscape of image descriptions, particularly regarding how appearance characteristics like race, gender, and disability are communicated. The researchers conducted interviews with 25 participants who use screen readers for most computing tasks and who identified with at least one marginalized identity along axes of race or gender. The study explores three interconnected areas: how participants negotiate their own and others' appearances nonvisually, their preferences for how appearance should be described in image descriptions, and their perspectives on AI-generated descriptions of appearance. The research is grounded in critical frameworks around disability, race, and gender, recognizing that these identity facets intersect and produce unique experiences of marginalization. Participants were recruited through connections in blind activist communities, specifically reaching out to BIPOC and LGBTQ+ groups. Interviews lasted 60-90 minutes and covered identity, social media use, appearance disclosure practices, language preferences, and reactions to AI-generated descriptions. The study fills an important gap in accessibility research by centering the perspectives of multiply marginalized people who are both the intended beneficiaries and potential targets of harm from image description technologies.
Key findings
Participants actively negotiated appearance through nonvisual strategies including asking friends, using AI tools like Seeing AI, reading post metadata, and interpreting voice and language cues. Nineteen of 25 participants had been misrepresented, with non-binary and transgender participants experiencing more negatively impactful misgendering than BIPOC participants experienced racial misrepresentation. Participants strongly preferred appearance language (skin tone, hairstyles, clothing) over identity language (race labels, gender labels, disability labels) in image descriptions when the describer does not know the photographed person's preferences. When describers know the person, participants wanted the person's own preferred language used. Participants identified six key contexts where appearance information was particularly valued: avatar creation, encountering unknown people, identity-related conversations, reading a room to find community, media representation, and seeking specific perspectives. Regarding AI-generated descriptions, participants were simultaneously excited about increased access and deeply concerned about bias, with some believing AI should not describe appearance at all given current limitations, while others favored imperfect descriptions over none.
Relevance
This paper is essential reading for anyone writing image description guidelines, building AI captioning systems, or creating alt text best practices. It challenges the common accessibility guideline to simply "describe what you see" by revealing the power dynamics and potential harms embedded in describing people's appearance. The distinction between appearance language and identity language offers a practical framework for image describers. For organizations, the findings suggest that image description policies should account for context, audience, and the photographed person's preferences rather than applying blanket rules. The study also raises critical questions about deploying AI for accessibility when those systems encode biases that disproportionately harm the very communities they aim to serve, cautioning against treating accessibility and fairness as separate concerns.
Tags: image descriptions · alt text · screen readers · visual impairments · race · gender · disability · AI bias · representation · intersectionality · identity
Standards referenced: WCAG