Are automatic captions good enough for accessibility rules?

Automatic captions can be a starting point, but nothing in the rules we read treats an unchecked machine draft as finished captions. The rules ask for captions that give the same access as the sound, and they do not name a recognition engine or a word-accuracy percentage that automatic output either meets or fails. Whether your automatic captions are good enough is decided by whether a viewer who relies on them gets the same information as a viewer who hears the video.

The ADA web rule for governments points to WCAG 2.1 Level AA

The Justice Department's ADA Title II web rule ties state and local government web content and mobile apps to WCAG 2.1 Level A and Level AA, unless compliance would be a fundamental alteration or an undue burden ¹. Within WCAG, captions for prerecorded video are a Level A requirement ², and captions for live audio are a Level AA requirement ³.

This page covers the Title II web rule. It does not cover the obligations of private businesses under other parts of the ADA, and we did not research those. A lawyer who knows your situation is the right source for your own obligations.

The rules ask for effective communication, not a named engine

Under the ADA's communication rules, covered entities must provide auxiliary aids and services when they are needed to communicate effectively with people who have communication disabilities ⁴. The rule does not say which tool to use. It asks whether the result communicates.

The FCC's quality rules for televised captions show one picture of what "communicates" means. Captioning shall provide non-speech information, such as the identity of speakers, music, sound effects and audience reaction, to the greatest extent possible ⁵. A machine draft that transcribes speech alone can miss all of that, even when most of the words are right.

Word accuracy alone does not settle whether captions are usable

Word error rate is the usual measure of an automatic engine, and it is a poor guide to usability on its own. In a 2017 study that proposed a new caption metric, word error rate correlated only 0.11 with ratings of caption usability from deaf and hard-of-hearing users ⁶.

The viewers themselves do not always rate automatic captions lower. In a 2026 online survey of 216 deaf and hard-of-hearing participants, live TV captions and automatic captions synchronized with the audio were rated similarly ⁷. That is a counterpoint, not a clearance, because it does not show that every automatic caption passes.

Automatic systems also invent text. A 2024 study of Whisper found that roughly 1 percent of audio transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ⁸. That study tested Whisper as it was in 2023, so a current engine may differ, but a reviewer still has to look for invented lines.

It depends on live or recorded video, the engine and the viewer

• Prerecorded video. Captions are required at Level A. A reviewed automatic draft can serve, provided the final file is accurate, synchronized and complete.
• Live audio. The requirement is Level AA, and live automatic captions are a trade-off. For a human service, see CART captioning. For the automatic trade-offs, see live captions.
• Your viewers. Test with deaf and hard-of-hearing viewers, not only with a word count. A caption can have a low error rate and still leave out a speaker change or a sound that matters.
• Vendor figures. Some vendor pages quote an accuracy benchmark. We did not find a regulation that sets a figure, so do not treat any vendor's number as the legal bar.

For how deaf and hard-of-hearing viewers judge AI captions, see whether AI captioning is accurate enough for deaf people.

What to do

1. Decide whether a rule applies to you. If you are a state or local government, the Title II web rule applies. For other organizations, ask a lawyer.
2. Treat automatic output as a draft. Review it against the audio from start to finish, not only the first minute.
3. Fix names, numbers and terms. Add speaker names where the speaker changes, and add sound tags for any sound the viewer needs.
4. Check the timing and line breaks. Play the video at the speed viewers will see it.
5. Test with viewers who depend on captions where you can, and keep a record of who reviewed the file and when.

For the full checklist, see the guide to captions and subtitles. The wider subject is in accessibility, and the question how accurate AI transcription is covers the accuracy side. For the format, see captions vs subtitles.