Three AI-specific problems
- Streaming output that re-announces on every token, making screen reader use unbearable
- Confidence shown only by colour, invisible to colour-blind users
- Dynamic regions that update without any announcement at all
None is hard to fix and all three are common, because they do not occur in conventional forms-and-tables software.
Getting streaming right
Use a polite live region and announce on completion rather than on every chunk. Sighted users get the streaming effect; screen reader users get a coherent sentence instead of a stutter.
Also provide a way to stop generation that is reachable by keyboard without leaving the output region.
Confidence needs more than colour
- A text label alongside any colour indicator
- An icon with a meaningful accessible name
- Contrast that meets the standard in both themes
- Sorting and filtering by confidence, not only visual scanning
Review interfaces must be keyboard-complete
This is both an accessibility requirement and the thing that makes review fast for everyone. Full keyboard operation, visible focus, logical order, and no action that requires a mouse.
It is the clearest case of accessibility work improving the product for all users rather than only for some.
Where the AI can help too
Generated alt text for images, plain-language summaries of dense documents, and readability checks on outgoing content. These are genuinely useful applications, provided a person confirms the alt text before it ships.
Automatically generated alt text that is wrong is worse than none, because it sounds authoritative.