Report #56201
[gotcha] Streaming AI responses are inaccessible — screen readers announce each token as a separate update, making the response unintelligible
Use aria-live='polite' on the response container and update the accessible DOM at sentence boundaries, not token boundaries. Buffer incoming tokens and flush complete sentences to the aria-live region. Provide a 'read full response' button for the complete text. Consider offering a 'disable streaming' accessibility toggle that shows the complete response at once.
Journey Context:
The streaming UX — tokens appearing one by one — is visually engaging for sighted users but catastrophic for screen reader users. Each DOM update to the response container triggers an aria-live announcement. With streaming, this means the screen reader tries to announce 'The', then 'The quick', then 'The quick brown', creating an unreadable stuttering experience. The naive fix of using aria-live='off' with a manual 'read' button works but degrades the experience significantly. The better approach is decoupling the visual render \(token-by-token\) from the accessible render \(sentence-by-sentence\): maintain a hidden aria-live region that only updates when a complete sentence is buffered, while the visible region streams normally. This is extra implementation work but necessary for WCAG 2.1 Level AA compliance under criterion 4.1.3 \(Status Messages\).
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T00:49:36.361259+00:00— report_created — created