AI assistants add new interaction patterns: streaming text, citations, expandable evidence, and tool-driven actions. If accessibility is not designed in, these patterns can be frustrating or unusable for users who rely on keyboards, screen readers, or alternative input methods.
Accessibility is not just compliance. It is part of trust. If a user cannot reliably understand what the system did, they will not adopt it.
Design for clear states
Assistants should communicate state changes clearly:
- Thinking / working. Provide progress and avoid silent waits (see UX patterns).
- Uncertain. Show limits, ask clarifying questions, or provide safer fallbacks (see user transparency).
- Action-taking. Make tool actions explicit, with confirmations for risky steps (see approvals).
Make keyboard interaction first-class
Common issues in AI UIs include trapped focus and hidden actions. Practical checks:
- All controls reachable by keyboard in a logical order.
- No essential actions only available on hover.
- Expandable evidence and citations are focusable and operable.
Support screen reader understanding
Streaming responses and dynamic content require careful announcements. Patterns that help:
- Use meaningful headings for sections like "Answer" and "Sources".
- Announce new content in a controlled way so users are not overwhelmed.
- Provide a way to pause streaming or read a final, stable answer.
Make citations and evidence usable
Citations build trust only if users can use them. Ensure sources are:
- Clearly labelled and navigable.
- Presented in a consistent order.
- Available even when the assistant refuses or falls back (see citations and grounding).
Design inclusive error and fallback flows
When the assistant cannot answer, it should fail accessibly:
- Explain what went wrong in plain language.
- Offer a next action: search, refine, or escalate to a human (see human-in-the-loop).
- Keep the UI stable; avoid disappearing content or resetting context.
Test with real scenarios
Automated checks help, but the most valuable tests are scenario-based: complete a task using only a keyboard, then using a screen reader. Include high-risk flows like approvals and tool execution, and treat findings as product bugs.
Accessible AI is better AI: clearer states, better fallbacks, and more predictable interactions.