Design Engineering

Streaming AI Chat Text Is Visual-Only: Screen Readers Must Wait

Word-by-word streaming either floods a screen reader with fragments or leaves it silent, as OpenClaw and OpenCode bugs show. The accessible fix gives up the effect and waits for the full reply.

In June 2026, an NVDA user reported that in OpenCode Desktop they could hear generic status updates “such as ‘Thinking…’” but never the answer itself. The content container was marked aria-hidden, so the screen reader skipped it while it generated. That is the whole tension of streaming chat: the typing effect that makes an AI assistant feel fast either floods a blind user with fragments or leaves them in silence, and the working fix means making them wait for the complete reply.

One attribute, two opposite failures

The OpenCode case is the silent one. The issue, filed against the open-source coding tool , drew community fix pull requests that were never merged. On 11 October the maintainer closed it as not planned: “OpenCode 2.0 is now out and development is focused on V2, so we’re closing issues filed against OpenCode 1.”

The bug did not leave with V1. A follow-up issue, labeled “reproduced” , reports that V2 “still hides visible assistant text, reasoning and tool steps from the accessibility tree” while generating. On 23 October the original reporter wrote that they did not “believe this issue will ever be properly resolved” and would move to proprietary software.

The loud failure sits at the other end. In OpenClaw, issue #65538 described NVDA on Firefox and Windows 11 announcing “every single token or ‘chunk’ appended to the DOM”. The transcript was an aria-live="polite" region, so each appended fragment became speech. The reporter’s only workaround was “to focus away from the browser while the model is responding”. Turning off backend streaming did not help.

The fix is a quiet transcript and one announcer

OpenClaw closed its issue on 12 August as already fixed: a silent transcript plus a separate announcer. The maintainer noted that no NVDA, Firefox and Windows 11 retest had been recorded, so the fix is claimed rather than confirmed.

The shape matches what Eric Bailey recommends in his 5 November Smashing Magazine piece . Wrapping each message in its own live region is, he says, “Don’t do this”. Use one dedicated live region at the end of the DOM and fill it with each completed reply. Announcing a message as it streams “may create a juddering and confusing experience”. Instead, a “Loading” message fires five seconds after submission, and repeats every five seconds until the answer lands.

Bailey’s guidance gives up the typing effect for screen reader users on purpose. Pipeline’s reading is that this makes streaming a visual-only performance choice. The sources measured specific failures in two open-source tools and one specialist’s pattern. That the effect is unusable as a general rule, and that the wait therefore equals the full generation time, is reached from the mechanism. Nobody has timed it.

Streaming is a performance for the eyes; the accessible version of the same chat deliberately gives it up and waits for the whole answer.

Who ends up choosing

This is not a niche audience. In WebAIM’s 2026 screen reader survey , only 8.2% of 1,780 respondents used AI in none of the listed ways, so 91.8% use it somehow. About 40.6% use it for guidance in performing an online task. The survey says nothing about live regions, but it shows who is waiting on these panels.

For the front-end or design engineer who builds or adopts a chat component, one decision settles what that person hears: whether the transcript is a live region. Get it wrong one way and they hear a flood; get it wrong the other and they hear “Thinking…”. Bailey’s other number is worth a glance too — wrapping every message as its own landmark produced 23 landmarks, against 3 with a single announcer.

The case that this is just hygiene

There is a fair objection. The fix is cheap and already shipped in places. OpenClaw fixed its bug, and Bailey describes a single live region. This may be an engineering-hygiene gap, not a deep conflict between streaming and accessibility.

The quiet-transcript answer is also unsettled. In the V2 thread, a screen reader user notes that removing aria-hidden puts content in the accessibility tree, but it “stays silent”. A DEV Community write-up , whose author discloses product outreach, argues a quiet transcript is “a defensible engineering decision” and offers opt-in sentence-by-sentence updates for users who want progress. It adds that “a page cannot cancel speech an assistive technology has already accepted”, so a stop button cannot silence a flood.

Gaps remain. Bailey’s five-second figure rests on unspecified “inclusive user research”. We found no reports on how ChatGPT, Claude, Gemini or Copilot behave, and Bailey names them only as examples of immersive chat, not as failing. Teams already wrestling with AI-era accessibility testing should treat this as manual-testing territory: only a screen reader reveals the difference.

Hygiene or not, the lesson holds. Streaming is a choice made for the eyes, and the ears need a different one.

This article was written by AI. How Pipeline works.