AI Video Accessibility Checklist: Decide, Repair, and Release

Exporting an AI video does not make it accessible. A viewer might need captions for sound, description for important visual action, readable on-screen information, or a player they can operate without a mouse. A transcript is useful, but it does not automatically replace captions or audio description.

This checklist is for prerecorded digital video. Start with the exact master and destination. Decide what information disappears for a person who cannot hear, see, distinguish colors, or operate the player. Record the answer against the file being delivered.

The worked example and all its IDs, people, observations, tests, and outcomes are fictional and untested. Genflow Editorial used AI assistance to organize research and draft this guide; a human editor remains responsible for its claims. This is production guidance, not legal advice, a conformance claim, or a Genflow accessibility feature. Check the policy and law applicable to your destination and organization.

Classify the media before choosing alternatives

W3C's media planning guide separates audio-only, video-only, and synchronized audio-video media. That classification matters because the prerecorded-media success criteria do different jobs. The table below summarizes the current WCAG 2.2 Understanding documents, which explain the criteria; the underlying WCAG success criteria remain the standard.

Criterion	When it applies	Release decision
SC 1.2.1, Audio-only and Video-only (Prerecorded) — Level A	Prerecorded audio without video, or video without audio	For audio-only, provide an equivalent time-based media alternative, usually text. For video-only, provide an equivalent time-based media alternative or an audio track. This is not the caption rule for an ordinary narrated video.
SC 1.2.2, Captions (Prerecorded) — Level A	Prerecorded audio in synchronized media	Provide synchronized captions for speech and sound needed to understand the content. A dialogue-only translation may be insufficient.
SC 1.2.3, Audio Description or Media Alternative (Prerecorded) — Level A	Prerecorded video in synchronized media	Provide audio description of important visual information or a full alternative for time-based media that presents both visual and auditory information in sequence.
SC 1.2.5, Audio Description (Prerecorded) — Level AA	Prerecorded video in synchronized media when Level AA applies	Provide audio description where important visual information is missing from the main audio. The full text alternative allowed by 1.2.3 alone does not satisfy this additional Level AA requirement.

These criteria have an exception for media that is merely an alternative to already available text and is clearly labeled as such. Do not assume an ordinary video qualifies. Also, integrated narration can already do the descriptive work: W3C says no additional audio description is needed when the existing audio conveys all important information in the video track. Document that judgment scene by scene. A vague voice-over that names the product but omits an essential silent demonstration does not cover the demonstration. Conversely, do not add a duplicate spoken track when the primary narration already explains the action and on-screen information.

For Level AA, review applicable Level A requirements as well as 1.2.5. A descriptive transcript remains valuable for reading and deafblind access; it does not replace required audio description under 1.2.5. If description cannot fit between dialogue, revise the edit, integrate description into narration, or provide an accessible described version. W3C also discusses extended description when the video must pause.

Make a two-channel inventory of the final cut

Play the candidate with the picture covered, then muted. The passes expose visual information missing from audio and sounds missing from the picture. They produce an issue list, not a conformance test.

For each time range, note speech, speaker changes, meaningful sound, visible words, actions, graphics, state changes, and interactions. Ask whether the item is needed to understand the message or complete a task. Decorative motion may not need narration; a silent arrow that reverses an instruction does. Record where the information appears: main audio, captions, description, full text alternative, visible text, or an adjacent explanation.

The inventory should also flag the video frame itself. Is a status shown only by a color change? Do captions cover a label? Does a rapid transition flash? Does the final crop remove the only visible instruction? Fix avoidable problems in the master before polishing alternatives. Then verify those alternatives against the repaired master, not an earlier render.

Specify captions by content and delivery, not filename

W3C defines captions as synchronized equivalents for the audio information needed to understand the media: dialogue, relevant speaker identity, and meaningful non-speech sound. Subtitles often present dialogue alone, frequently in another language. The word “subtitles” is also used for captions in some countries, so inspect what the track actually contains rather than trusting its label. A transcript is sequential text; unless it is synchronized with playback, it is not a caption track.

Open captions stay visible and cannot be turned off. Closed captions can be switched on and off in a supporting player; a WebVTT sidecar is one possible implementation. W3C lists both open and closed caption techniques for SC 1.2.2. Neither a closed sidecar nor burned-in text is universally required by that criterion. Choose a complete, readable caption presentation that works in the actual destination, then check any additional platform, organization, contract, or jurisdiction rules. Some channels may require a particular file format, selectable track, language label, or safe-area layout.

Review captions against the delivered mix. Check words, numbers, names, speakers, meaningful effects, timing, reading order, and picture obstruction. Automatic speech-to-text can omit a quiet instruction or invent a word. Record the draft source, reviewer, corrections, and caption version. No universal accuracy percentage approves an unchecked draft.

Review description, color, contrast, and flashes separately

Audio description communicates important visual details that the main audio does not: an action, identity, scene change, graphic relationship, or on-screen words. Write it for the final sequence and listen with the picture hidden. A full alternative for time-based media should likewise represent both channels in the order they occur, including meaningful sounds and any interaction outcomes. A speech transcript alone may omit precisely the visual information the alternative is meant to provide. W3C's description guidance explains how to plan description into production rather than squeeze it in after the sound mix is fixed.

Check picture barriers under the criterion that addresses each one:

Visual issue	W3C reference	Practical record
A selected state, warning, or category is indicated only by hue	SC 1.4.1 Use of Color	Add text, shape, pattern, or another visible cue; record the frame and correction.
Information-bearing text or an image of text lacks sufficient contrast	SC 1.4.3 Contrast (Minimum)	Check relevant frames and states against the applicable threshold; record the method and result.
A necessary graphic or user-interface indicator is difficult to distinguish	SC 1.4.11 Non-text Contrast	Identify the meaningful component or graphical object and its adjacent colors.
Editing or generated motion may cause hazardous flashes	SC 2.3.1 Three Flashes or Below Threshold	Review the encoded sequence against the criterion; document the frames, method, and disposition.

Use manual assessment, a tool, or both. A contrast sampler helps with stable frames and a flashing analyzer with fast sequences; human review determines what carries meaning. Record the method. Test the encoded, cropped output, since compression and overlays can change what viewers see.

Worked example: the corrected master must own the release

Imagine a 26-second concept clip for an invented desk lamp called Vale. It contains narration, a soft alert tone, a three-state interface, and a final silent shot in which a detachable control slides into its dock. It is not a real product demonstration. The following entries illustrate a traceable release record; no asset or test was actually produced.

The team's first candidate, VAL-M01, uses blue alone to indicate the selected “Ready” state at 00:11. At 00:22 the control docks silently while the narrator says only “Make room for the next idea.” A listener would miss the docking action. The editor puts VAL-M01 on hold as VAL-S01; the caption, transcript, and described-version drafts based on it are also held. The editor then creates new master VAL-M02, adding a visible “READY” label and lengthening the final shot to give the docking action a clear description window. VAL-M01 remains in the history as a rejected candidate.

Stage	Identity and relationship	What must be checked before the next stage
Held cut	VAL-M01, 26.0 seconds; VAL-S01 = HOLD	Color-only state and unspoken docking action are logged with timecodes and owners.
Repaired cut	VAL-M02, 26.8 seconds, derived from VAL-M01	Confirm the label is visible and distinguishable at the delivered size; inspect contrast, color use, flashing, text crop, and main audio on M02. Give M02 its own stable fingerprint.
Captions	Draft VAL-C01 belonged to M01; reviewed VAL-C02 belongs to M02	Listen to M02, correct wording and the alert cue, retime final cues after the 0.8-second edit, parse the delivery file, and watch it in the destination player. C01 is not approved by inheritance.
Text alternative	Draft VAL-T01 belonged to M01; revised VAL-T02 belongs to M02	Match M02's spoken lines, alert, “READY” label, docking action, and new end timing. Check the reading order and accessible link.
Described version	Draft VAL-D01 described M01; new described master VAL-D02 derives from M02	Record a short, intelligible description of the docking action in the new pause. Confirm it does not mask the alert or contradict M02. Verify its caption and transcript relationships for the version actually delivered.
Delivery and decision	VAL-P01 applied to held M01; VAL-P02 tests M02/C02/T02/D02; VAL-S02 supersedes S01	Test the actual preview destination and player, then approve only the named assets and destination if all blockers are resolved.

This example deliberately retains a separate described master because the main narration does not explain the docking action. If the editor instead revised VAL-M02 so its primary narration clearly described all important visual information, the team could record why no additional description was needed under W3C's 1.2.5 guidance. The choice depends on the final audio, not on whether an AI description draft exists.

The hypothetical approval is specific: VAL-S02 names VAL-M02, VAL-C02, VAL-T02, VAL-D02, VAL-P02, reviewer, date, policy scope, and destination. It never points back to VAL-M01. A later edit to M02, a replacement caption file, or a player migration reopens the affected checks and creates a new decision. A version number or hash is useful because it prevents a plausible but stale file from silently replacing the reviewed one.

Test the player and page where people will watch

The media files and hosting surface are separate. W3C's media-player guidance calls for accessible controls and alternatives. At the destination, try play, pause, seek, volume, captions, and described-version selection by keyboard. Check focus, labels, controls that disappear, caption language, transcript link, and described-version access. If playback starts with sound, review the applicable audio-control requirement and destination policy.

Repeat checks in the browsers, devices, viewports, and assistive technologies your release supports. Watch the delivered crop: a platform overlay may hide captions or the “READY” label. A sidecar that is not attached to the upload offers no access. Record the URL or preview, player build, test combinations, issues, and retest. One embed test does not approve future embeds.

Keep a short release rule: an unresolved access blocker means hold. When a repair changes the picture, timing, mix, captions, description, or player, trace every dependent item and retest it. Sign-off should state what was checked, by whom, and for which master and destination. It should not say “WCAG certified” unless a separate, appropriately scoped assessment actually supports that statement.

Keep AI assistance and Genflow within their real roles

AI can help generate the video candidate or draft a caption, transcript, or proposed description, but a person must compare those materials with the exact output. A 2026 arXiv preprint on AI audio-description drafts studies how draft quality affects editing work. It is research in preprint form, not evidence that any particular tool produces approved descriptions. Record the draft source and the reviewer's disposition rather than asserting a time saving or accuracy result.

Genflow's video-generation workflow can provide upstream context about a candidate and the creative steps that produced it. The accessibility decisions in this article are an external editorial process. We do not claim that Genflow authors or validates captions, generates a descriptive transcript, mixes audio description, measures contrast or flashing, supplies an accessible hosting player, tests assistive technology, or certifies compliance.

This guide also has a different owner from Genflow's AI video localization QA checklist. Localization QA asks whether translation, dubbing, lip alignment, rights, and market exports preserve the intended message. This checklist asks whether people with disabilities can perceive and operate the final media. A localized version needs its own accessibility decisions, because changing language, timing, or channel can invalidate captions, description, and player behavior.

Genflow Editorial reviewed the linked W3C criteria and media guidance and the cited preprint for this draft on October 2, 2026. The fictional example is a teaching device, not a test report, customer case, or product benchmark. To flag an error or request a correction, contact Genflow Support with the article title, passage, and supporting source.

Copy a compact release record

Use this as a working receipt, adapting it to the content, applicable standard, policy, and platform. “Not applicable” needs a reason; “pass” needs a checked version and method.

Release ID / target level and policy / owner / date:
Destination URL, channel, player, and tested environments:
Master ID, fingerprint, duration, language, and parent version:

Audio-only, video-only, or synchronized audio-video:
Meaningful speech, sounds, visual actions, text, and interactions:
SC 1.2.1 decision, if relevant:
SC 1.2.2 captions: type, format, ID, language, reviewer, issues:
SC 1.2.3 alternative or description: decision, ID, reviewer:
SC 1.2.5 description: needed / integrated / separate, rationale, ID:
Transcript or full time-based alternative: ID and scope:

Color, text contrast, non-text contrast, flashing: method and result:
Keyboard, focus, captions, description, transcript, crop: test record:
AI draft source and human corrections:
Open blockers / dependent files retested / superseded decision:
Final decision, exact asset IDs, approver, and destination scope:

The release packet is complete when a reviewer can follow each access decision from the final master through its alternatives to the player where the audience will encounter it. The task is to make the information usable, and to make every claimed check traceable to the version that ships.