AI products inherit every accessibility obligation of conventional software and add uncertainty, streaming updates, generated structure, and variable content. A compliant shell cannot compensate for an inaccessible response.
Input and control
- Every field and control has an accessible name and visible focus state.
- All actions work with a keyboard and do not depend on drag, hover, or timing.
- Voice is optional; equivalent text input and output are available.
- Examples and guidance explain what the system can accept.
Generated output
- Headings, lists, tables, citations, and code use semantic structure.
- Streaming updates do not repeatedly interrupt screen-reader users.
- Status changes are announced at the appropriate priority.
- Generated images have useful alternatives or are clearly marked when descriptions are unavailable.
- Color is not the only signal for confidence, error, or provenance.
Accessibility must be evaluated against the range of content the system can create, not only the ideal example in the design file.
Error and recovery
- Errors explain what happened and how to continue.
- Users can pause, retry, edit, undo, and restore prior work.
- Focus moves predictably after generation, validation, and modal actions.
- Time limits can be extended or removed.
Cognitive accessibility
Use plain language, stable navigation, manageable choices, and summaries for long outputs. Allow users to control verbosity and preserve context across steps. Avoid presenting uncertain generated content with the visual authority of a verified fact.
Test with real variation
Combine automated checks, keyboard testing, screen readers, zoom, high contrast, reduced motion, and people with disabilities. Include long answers, empty states, unsafe requests, malformed content, slow generation, and failures. Accessibility is complete only when the product remains usable outside its happy path.