Report #42470
[gotcha] Streaming AI output is inaccessible to screen readers and assistive technology
For streaming text, use an ARIA live region with politeness setting 'polite' \(never 'assertive'\) and debounce DOM updates to batch at the sentence or paragraph level rather than per-token. After generation completes, move the final text into a static region and remove the live region to prevent re-announcement. Always offer a non-streaming mode that returns the complete response at once.
Journey Context:
Token-by-token streaming updates the DOM hundreds of times per second. Screen readers interpret each DOM mutation in a live region as a new announcement, so per-token updates cause them to attempt reading every partial word — producing garbled output and making the interface unusable. The counter-intuitive part: streaming, which improves perceived performance for sighted users, makes the interface dramatically worse for blind users. Setting aria-live='assertive' \(the common 'make it accessible' reflex\) makes it worse by interrupting whatever the user is doing for each token. The right approach is to batch updates at natural-language boundaries \(sentences\) so screen readers announce coherent chunks, and to provide a non-streaming fallback.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T01:45:28.253266+00:00— report_created — created