Foundations
What is AI accessibility?
A practical framework for assessing whether people can access, understand, evaluate and remain in control of AI systems.
AI accessibility covers more than an interface
AI accessibility concerns whether disabled and neurodivergent people can use an AI system, benefit from it and remain protected when it influences a service or decision. Keyboard access, captions and screen-reader compatibility matter, but they are only the starting point.
AI also introduces cognitive demands: interpreting probabilistic language, recognising uncertainty, remembering context across a conversation and deciding when an answer should not be trusted. A technically operable product can therefore remain inaccessible in practice.
Three different accessibility questions
Accessibility of AI asks whether people can operate and understand the system itself. Accessibility through AI asks whether an AI feature removes a barrier, such as by producing captions or simplifying information. Accessibility under AI asks what happens to a person when an automated system recommends, ranks or decides.
Teams should examine all three questions separately. An assistive feature does not compensate for an inaccessible interface, and an accessible interface does not make an unfair or unchallengeable automated decision acceptable.
A practical capability test
A useful review asks whether a person can access the interaction, understand the output, evaluate its evidence and uncertainty, make an informed choice, retain meaningful control and recover after an error. These capabilities turn an abstract commitment into questions that can be tested with people using real tasks.
For example, a recruitment tool should not only announce controls correctly. A candidate must also understand when AI is being used, what information matters, how to correct inaccurate data and how to reach a person if the process fails.
Evidence to request
Claims such as accessible, inclusive or human-centred are not evidence by themselves. Buyers and governance teams should request test scope, participant characteristics, known limitations, failure reports, escalation routes and evidence that changes were made after disabled people identified barriers.
Evidence should cover the complete journey rather than a demonstration of one successful feature. It should also be reviewed after material model, interface or policy changes because accessibility can regress when an AI system changes.
Where to begin
Start with the decisions and tasks that carry the greatest consequences. Map who may encounter barriers, include disabled and neurodivergent people in research, and test comprehension and recovery as deliberately as technical conformance.
W3C cognitive accessibility guidance, WCAG, the NIST AI Risk Management Framework and applicable law provide useful foundations. They do not replace direct evidence from the people expected to use or be affected by the system.