Screen-reader contracts for AI streaming, citations, errors

Trust is a contract, not a feature

When an AI answer streams, a screen reader can sit silent. The user assumes the page is frozen, not that the model is generating. A trustworthy AI feature gives every state a non-visual contract: what changed, what it means, and what the user can do next.

The pattern fails when the interface assumes sight, hearing, or a steady pointer. The redesign is a set of explicit states with announcements, visible alternatives, and recoverable failures.

The platform is already setting the bar

Apple announced AI-powered accessibility capabilities for VoiceOver, its screen reader, Magnifier, its low-vision tool, Voice Control, and Reader in honor of Global Accessibility Awareness Day on May 21. VoiceOver Image Explorer will provide deeper systemwide image descriptions for photos, scanned bills, and personal documents. AI-generated subtitles will be available for uncaptioned video on iPhone, iPad, Mac, Apple TV, and Apple Vision Pro. Voice Control's natural-language input will be available in English for the U.S., Canada, the UK, and Australia.

Vision Pro wheelchair control will begin by supporting Tolt and LUCI drive systems in the U.S. Apple's name-recognition feature supports 50 languages. Magnifier will get a high-contrast interface aimed at users with low vision.

Those moves make the expectation concrete. If a system can describe an image, caption a video, or accept natural-language commands, a product team cannot hide behind the excuse that the AI output is too fluid to make accessible.

Screen readers convert on-screen text into spoken output for blind users. Deaf or hard-of-hearing users can use captions for web content. Voice-recognition software lets users with limited mouse control operate computers and devices through spoken commands. Image alt text communicates the purpose of pictures, illustrations, and charts to users who cannot see them.

The redesign is a stable state

Wrap model output in a component that owns status, meaning, and action. The model can be slow, wrong, or silent. The component must still show the state, name the source, and offer a next step. Spec the states before the model responds, build the component around them, and review the live region in the same pass.

That component is the part of the UI that can be relied on. It should survive a failed request, a partial answer, a long pause, and a user who cannot see the screen. When the model changes, the component stays stable.

A demo shows the model working; a feature supports the user when the model does not.

The checklist is the release check

For each AI output, define the screen-reader announcement, low-vision contrast and layout, caption/subtitle behavior, alt text, citation access, voice option, and error recovery. Do it in spec, build, and review before release. Run each test with the screen reader active and the pointer disabled. A state that cannot be understood without sight, hearing, or a pointer is not done.

• Streaming: announce start, pause, and completion; test by muting audio and reading the live region that announces changes.
• Citations: expose each source as a named, focusable link or button; test by tabbing to it without a mouse.
• Voice: provide a text transcript and a way to stop, replay, or replace audio; test with sound off.
• Captions: keep AI captions and subtitles synchronized, editable, and visible; test with audio muted and a low-vision zoom.
• Alt text: describe purpose, not pixels; test by reading the alternative text alone.
• Errors: announce the problem, the affected field, and the next action; test by submitting an invalid form.
• Low vision: preserve contrast, spacing, and target size; test at a larger zoom.
• Recovery: offer retry, copy, report, or fallback; test after a timeout or failed generation.

A silent token stream can look like a frozen page, and a source hidden in a hover tooltip is invisible to keyboard and screen-reader users. Audio alone does not replace readable text. Captions carry time-based speech; alt text carries image purpose.

Contrast and spacing are part of the contract. Failed generations and timeouts both need a recovery path. Form error messages should be announced automatically to screen-reader users when a field is entered incorrectly. Screen readers do not announce text color, so color-only cues can hide information from blind users. WCAG and Section 508 are existing technical standards that the U.S. Department of Justice identifies as helpful guidance for accessible website features.