Helpful reading
If you’re newer to AI and accessibility, we’ve included links to a few talks and readings, grouped by topic, that you could find helpful.
What an AI agent is
An AI agent is software that acts for a person, for example by reading a page, filling in a form or changing how an app looks. An agent for accessibility does those things to remove the barriers a person runs into, the way that person asks.
- What are AI Agents?, a video from IBM Technology (2024, about 12 minutes). Watch this for a whiteboard talk on how an AI system looks things up and calls on other programs to finish a task.
- Large Language Models explained briefly, a video from 3Blue1Brown, made for the Computer History Museum (2024, about 8 minutes). A short animation on how the language models inside most AI agents work.
- What are AI agents?, from IBM. A written definition, with plain examples of agents planning steps and using tools.
- Building effective agents, from Anthropic (2024). For developers. It separates systems that follow fixed steps from ones that choose their own next step.
- LLM Powered Autonomous Agents, by Lilian Weng (2023). For developers. It sets out an agent’s parts: a language model at the center, with planning, memory and tools around it.
- On the Opportunities and Risks of Foundation Models, by Bommasani and colleagues (Stanford, 2021). Experts from many fields take a broad look at the large AI models agents are built on.
- Building a general-purpose accessibility agent, and what we learned in the process, from the GitHub blog. The team behind the accessibility features in GitHub Copilot on how they built them and what they learned.
- A11y-CUA Dataset: Characterizing the Accessibility Gap in Computer Use Agents, by Mohanbabu and colleagues (CHI 2026). It shows what happens when blind and low-vision people use agents that operate a computer for them.
What an AI agent could do for someone
Whether someone sees, hears, moves or thinks differently, an agent can change the interface to fit them, instead of asking them to fit the interface.
- Improving the performance of motor-impaired users with automatically-generated, ability-based interfaces, by Gajos, Wobbrock and Weld (CHI 2008). SUPPLE, an early system, built interfaces to fit each person’s motor abilities.
- Interaction Proxies for Runtime Repair and Enhancement of Mobile Application Accessibility, by Zhang and colleagues (CHI 2017). It fixes an app’s accessibility from the outside, without touching the app’s own code. The authors have since come to dislike the word “proxies”, given the history of the term.
- Accessibility Came by Accident: Use of Voice-Controlled Intelligent Personal Assistants by People with Disabilities, by Pradhan, Mehta and Findlater (CHI 2018). How disabled people use voice assistants that nobody designed with them in mind.
- Use of Intelligent Voice Assistants by Older Adults with Low Technology Use, by Pradhan, Lazar and Findlater (ACM Transactions on Computer-Human Interaction, 2020). Older adults who rarely use technology tried voice assistants at home, and the paper follows how they took to them.
- An Autoethnographic Case Study of Generative Artificial Intelligence’s Utility for Accessibility, by Glazko and colleagues (ASSETS 2023). Researchers tested generative AI against their own access needs and recorded where it helped and where it fell short.
- Autoethnographic Insights from Neurodivergent GAI “Power Users”, by Glazko and colleagues (CHI 2025). Neurodivergent people who rely on generative AI every day explain how they make it work for them.
- Sign Language Recognition, Generation, and Translation: An Interdisciplinary Perspective, by Bragg and colleagues (ASSETS 2019). Deaf and hearing experts set out what sign language technology has to get right before it really serves Deaf people.
- Evaluating the Usability of Automatically Generated Captions for People who are Deaf or Hard of Hearing, by Kafle and Huenerfauth (ASSETS 2017). It shows that counting caption errors misses which mistakes actually confuse Deaf readers.
- SoundWatch, by Jain and colleagues (ASSETS 2020). Deaf and hard of hearing people tested a smartwatch that names sounds like sirens or doorbells.
- Making Content Usable for People with Cognitive and Learning Disabilities, from the W3C (2021). The main web standards guide to making content clear and usable for people with cognitive and learning disabilities.
- Digital Comprehensibility Assessment of Simplified Texts among Persons with Intellectual Disabilities, by Säuberli and colleagues (CHI 2024). It tests whether simplified texts really help readers with intellectual disabilities, and finds the benefit is modest.
- Shared Stories, Shared Bonds: People with Dementia Exploring Generative AI Together, by Arets and colleagues (CHI 2026). People living with dementia try generative AI tools together, as something they do as a group.
- GPII and Cloud4all. Earlier projects on preferences that go with a person from one device or service to the next.
- ISO/IEC 24751. An international standard for describing a person’s needs and preferences so systems can adapt to them.
- Preference signals browsers already support, such as the setting that asks websites to reduce motion, defined in the W3C’s Media Queries Level 5. A small example, already working, of a person’s preference reaching every site they visit.
The potential harms
Agents can fail in ways that hurt most the people they’re meant to help. They can keep acting after a person asks them to stop, describe what a person can’t do rather than what they ask for, personalize in harmful ways, or choose fixes for a person rather than letting them choose, as many accessibility overlays do.
- Race, Disability, and Technology: A Call to Action for Accessibility Researchers, by Desai and colleagues (ACM Transactions on Accessible Computing, 2026). It asks researchers to look at where race and disability meet in the technology we build and study.
- I Don’t Trust it, but I Use it: Navigating Trust, Privacy, and Identity in Disabled People’s Use of Generative AI, by Johnson and colleagues (CHI 2026). How disabled people weigh trust, privacy and identity when they use generative AI anyway.
- Minor Resistance: The Everyday Politics and Power Dynamics of Assistive Technology Adoption, by Hsueh and colleagues (ASSETS 2025). Power and everyday politics shape whether people take up assistive technology.
- Systemic Biases in Sign Language AI Research: A Deaf-Led Call to Reevaluate Research Agendas, by Desai and colleagues (2024). Deaf researchers reviewed 101 papers and found that sign language AI often ignores what Deaf communities want.
- WFD and WASLI Statement on Use of Signing Avatars, from the World Federation of the Deaf and the World Association of Sign Language Interpreters (2018). The case against replacing qualified human interpreters with computer signing.
- AI and the Future of Accessible Work, by Piper and colleagues (Communications of the ACM, 2026). What AI could change about work for disabled people.
- Overlay Fact Sheet. Accessibility practitioners and disabled people signed this open statement on why overlays don’t make a site accessible.
Designing AI agents with accessibility at the front
Design with the disabled people you’re designing for, from the start. Don’t only test with them at the end.
- Ability-Based Design: Concept, Principles and Examples, by Wobbrock and colleagues (ACM Transactions on Accessible Computing, 2011). The founding paper on designing for what people can do. There are many more, but start here.
- Is There Design-for-All?, by Simon Harper (Universal Access in the Information Society, 2007). An argument for design for one: interfaces built around each person’s needs, which agents now make possible.
- Design for User Empowerment, by Richard Ladner (ACM Interactions, 2015). The case for disabled people creating solutions for themselves and for others.
- Exploring Disability Culture Through Accounts of Disabled Innovators of Accessibility Technology, by Desai, Mankoff and Ladner (ASSETS 2025). Disabled people who built their own accessibility technology tell the story in their own words.
- Participation Without Power? A Systematic Review and Call to Center Shared Authority in Accessibility Research, by Jang and colleagues (ASSETS 2026). Much co-design work doesn’t let disabled people lead at most stages of design, and papers with a disabled author tend to go further.
- Anticipate and Adjust: Cultivating Access in Human-Centered Methods, by Mack and colleagues (CHI 2022). How to make research and design sessions accessible to the people taking part in them.
- UDL Guidelines, from CAST (version 3.0, 2024). A framework for designing so that people who learn differently are included from the start.
- Web Content Accessibility Guidelines (WCAG) 2.2, from the W3C. The standard most accessibility testing checks against.
- Planning and managing web accessibility, from the W3C Web Accessibility Initiative. Where a team should start if it wants to build accessibly from the beginning.
- The WebAIM Million. Now in its eighth year (2026), WebAIM’s yearly report on the accessibility of the top million home pages shows how far the web still has to go.
- ScreenAudit: Detecting Screen Reader Accessibility Errors in Mobile Apps Using Large Language Models, by Zhong and colleagues (CHI 2025). AI finds the problems screen reader users would run into in mobile apps.
- Accessibility review agents, from the Community Access project on GitHub. Each agent reviews one area of accessibility, such as ARIA, contrast, keyboard navigation, forms or tables.
Think something’s missing from this page?
Suggest a talk or reading through our contact form.