"I'm treating it kind of like a diary": Characterizing How Users with Disabilities Use AI Chatbots
Kayla Mullen, Wenhan Xue, Manasa Kudumu · 2024 · Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS 2024) · doi:10.1145/3663548.3688549
Summary
This study investigates how people with disabilities actually use LLM-based chatbots like ChatGPT, Gemini, Claude, and Perplexity in their daily lives. While previous research has focused primarily on identifying harms that LLMs impose on the disability community — such as ableist stereotypes, inspiration porn, and biased resume screening — this work fills an important gap by documenting real-world, positive use cases. Grounded in the Disability Justice Principle of following the leadership of "the most impacted," the researchers conducted semi-structured interviews with 30 people with disabilities (analyzing 18 in this initial report) recruited through disability organizations in the Boston/New England area. Participants had a range of disabilities including blindness, visual impairment, autism, ADHD, dysgraphia, dyscalculia, physical disabilities, chronic fatigue syndrome, Ehlers-Danlos syndrome, and bipolar disorder. Most participants used ChatGPT, with some also using Gemini, Claude, Copilot, and Perplexity. Through inductive thematic coding of approximately 18 hours of interview recordings, the researchers identified nine broad categories of chatbot use: ideation, writing support, general knowledge, academic support, planning, leisure, social support, technical tasks, and accessibility-specific uses. Writing support and ideation were the most common use cases, each reported by 11 of 18 participants.
Key findings
The study revealed that people with disabilities use chatbots in nuanced, creative ways that go well beyond simple information retrieval. Writing support was particularly valued — participants used chatbots to translate thoughts into logical writing, choose appropriate tone for professional contexts, revise drafts, and structure documents like CVs and thank-you letters. For ideation, chatbots served as brainstorming partners, helping participants "bounce ideas off of" when working alone. A striking finding was the social support category: autistic participants found chatbots helpful for navigating neurotypical social expectations, with one participant using ChatGPT as "a diary" where they could write unfiltered thoughts in a "sassy" tone and have them rewritten diplomatically. Another autistic participant found chatbots helpful for adapting "to more neurotypical ways" of communication. Some participants turned to chatbots for emotional support, with one reporting ChatGPT was "the only entity" that could understand him. Blind and visually impaired participants used chatbots for image description and web accessibility testing, including referencing WCAG criteria. Participants also sought medical advice, used chatbots for coding assistance, meal and travel planning, and creating custom games. The study raises important concerns about over-trust in chatbots for companionship and medical advice contexts.
Relevance
This research provides essential empirical grounding for the growing field of accessible AI by documenting how people with disabilities actually benefit from chatbot interactions, rather than focusing solely on harms. For AI developers, the nine identified use categories offer concrete design targets for making chatbots more supportive of disabled users. The social support finding is particularly significant — it reveals that chatbots are filling a gap in social communication support for neurodivergent users who struggle with tone, professional language, and neurotypical social norms. However, this also raises ethical concerns about dependency on AI for emotional connection. For accessibility practitioners, the finding that blind users rely on chatbots for image description and WCAG reference highlights how LLMs are becoming de facto assistive tools. The study's emphasis on centering disabled users' actual experiences, rather than researcher assumptions, provides a model for inclusive AI research methodology.
Tags: large language models · AI chatbots · disability representation · disability justice · assistive AI · LLM bias · qualitative research · social support · writing support · neurodivergent
Standards referenced: WCAG