"I look at it as the king of knowledge": How Blind People Use and Understand Generative AI Tools
Rudaiba Adnin, Maitraye Das · 2024 · ASSETS 2024: 26th International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3663548.3675631
Summary
This study investigates how blind individuals use and make sense of mainstream generative AI tools, based on semi-structured interviews with 19 blind participants who regularly use tools like ChatGPT, Google Gemini, Microsoft Copilot, Claude, and Be My AI (the GenAI-powered image description feature of Be My Eyes). Conducted between January and March 2024, participants demonstrated their workflows via screen sharing on Zoom, showing chat histories and live interactions. The research addresses three questions: how blind people use GenAI tools, how they navigate challenges and biases, and what mental models they develop. Participants used GenAI for diverse content creation tasks (copywriting, emails, resumes, cover letters, creative writing, programming) and information retrieval (product research, planning events, health information). A distinctive use case was visual question answering through Be My AI, which participants valued for providing richer, more systematic image descriptions than sighted helpers typically offer. However, participants encountered significant accessibility barriers: unlabeled buttons for copying, regenerating, and voting on responses; no heading labels or landmarks for screen reader navigation; no notification when response generation completed; and extraneous ads and sample prompts cluttering interfaces. Participants developed elaborate workarounds, such as repurposing the graphics navigation shortcut to jump between prompts by exploiting the fact that user avatars were announced as "graphic" by screen readers.
Key findings
Participants developed seven distinct mental models of GenAI: "Google Search on Steroids" (GenAI pulls information from the internet), "King of Knowledge" (GenAI queries a massive database containing all information), "Word Generating Machine" (GenAI predicts next words based on language patterns), "Stores and Reuses User Prompts" (GenAI learns from other users' conversations), "More In-depth AI" (GenAI is a superior version of voice assistants), "Partner, Friend, Mentor, Secretary" (anthropomorphized roles), and "Still a Computer, Not a Human" (recognizing fundamental limitations). Many of these models were flawed or oversimplified, leading to problematic trust patterns—for example, the belief that GenAI always retrieves accurate factual information. Blind participants' mental models were uniquely shaped by the inaccessibility of GenAI interfaces; many were unaware of upvote/downvote buttons because they were unlabeled, leading them to provide feedback through chat prompts instead and incorrectly assuming these corrections would improve future responses. Participants identified four factors governing their verification decisions: context of use, stakes involved, verifiability of the information, and believability of the response. GenAI tools produced ableist content—describing blind characters as "courageous" and "resilient" in stories, expressing pity when asked about blindness-related tasks, and making gendered assumptions in image descriptions. Participants expended significant "advocacy labor" correcting these biases in hopes of improving future outputs.
Relevance
This paper is essential reading for anyone designing or deploying GenAI tools with accessibility in mind. It demonstrates that surface-level text accessibility is insufficient—unlabeled buttons and poor UI structure create divergent mental models between blind and sighted users, with potentially serious consequences for trust and information verification. The four-factor verification framework (context, stakes, verifiability, believability) provides a useful lens for understanding when blind users are most vulnerable to GenAI misinformation. The finding that blind users tend to trust Be My AI image descriptions more than sighted help because they are more detailed highlights a double-edged sword: richer descriptions improve access but also increase the believability of hallucinated content. The paper calls for GenAI interfaces to build in accessible verification mechanisms rather than requiring users to switch between multiple apps. It also raises a critical policy tension: measures to combat GenAI harms (like removing face descriptions for privacy) can simultaneously revoke accessibility benefits for blind users.
Tags: generative AI · blindness · visual impairment · ChatGPT · mental models · AI accessibility · screen readers · information verification · ableism in AI
Standards referenced: WCAG