An Autoethnographic Case Study of Generative Artificial Intelligence's Utility for Accessibility
Kate S. Glazko, Momona Yamagami, Aashaka Desai, Kelly Avery Mack, Venkatesh Potluri, Xuhai Xu, Jennifer Mankoff · 2023 · Proceedings of the 25th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS '23) · doi:10.1145/3597638.3614548
Summary
This paper presents an autoethnographic case study in which seven accessibility researchers—both with and without disabilities—explored generative artificial intelligence (GAI) tools over a three-month period to assess their potential for supporting accessibility needs. The team used a range of GAI tools including ChatGPT, GPT-4, DALL-E 2, Midjourney, GitHub Copilot, and Bing Chat across their daily academic and personal tasks. The study employed a collaborative autoethnographic methodology where researchers documented their GAI interactions through structured diary entries, then collectively analyzed these experiences through affinity diagramming and thematic analysis. The researchers identified two broad categories of GAI use: employing GAI to meet one's own access needs, and using GAI to make content accessible for others. The paper presents seven detailed vignettes illustrating these uses, including using language models for summarization and information extraction to manage brain fog, facilitating interpersonal communication for an autistic researcher, generating visual imagery for someone with aphantasia and motor impairments, assisting a blind developer with GUI and visualization design, converting LaTeX tables into accessible formats, improving slide accessibility, and adjusting visualization colors for color vision deficiency. The research provides a nuanced, first-person perspective on the practical realities of using GAI tools for accessibility, going beyond theoretical potential to document actual experiences of disabled researchers navigating these emerging technologies.
Key findings
GAI tools showed genuine promise for accessibility but with significant caveats. The tools were most useful for low-stakes, easily verifiable tasks—such as generating initial drafts, suggesting color palettes, or extracting information from dense text—where users could readily check outputs against their own knowledge. A critical finding was the "verifiability paradox": GAI was least reliable precisely when verification required the same accommodation the tool was providing, such as a blind developer needing to verify visual output. The researchers identified several concerning patterns including built-in ableism, where GAI produced non-representative disability imagery (e.g., depicting disabled people only in medical or pitying contexts) and subtly biased language in communication assistance. GAI tools also exhibited "false promises," confidently providing incorrect accessibility solutions—for example, generating ARIA code that appeared valid but contained errors that would degrade rather than improve accessibility. The study found that GAI currently functions best as a starting point or brainstorming aid rather than an end-to-end accessibility solution, and that the gap between GAI's confident presentation and actual accuracy poses particular risks for users who lack the expertise to identify errors in accessibility-related outputs.
Relevance
This study offers essential ground-truth evidence for accessibility practitioners considering GAI integration into their workflows. The finding that GAI tools can confidently produce incorrect accessibility code—particularly ARIA implementations—is a critical warning for developers who might over-rely on these tools for compliance work. The verifiability paradox has direct implications for assistive technology design: tools that cannot be verified by their intended users require additional safeguards or human-in-the-loop processes. The documented instances of ableist bias in GAI outputs highlight the need for disability-inclusive training data and evaluation frameworks. For organizations, this research suggests that GAI can augment accessibility workflows but should not replace expert review, and that disabled users' experiences with these tools must inform their development. The autoethnographic methodology itself offers a model for centering disabled perspectives in technology evaluation.
Tags: generative AI · accessibility · autoethnography · assistive technology · disability · ChatGPT · large language models
Standards referenced: WCAG · ARIA