Did your AI captions miss the sound that explains the scene?
Imagine a character stopping mid-sentence because someone knocks. Leave the knock out of the captions and the pause can look like a change of heart. You've changed the scene without mistranscribing a word. W3C's existing accessibility guide says captions include the non-speech audio needed to understand the content, not just dialogue. It also says automatic captions need checking for full accuracy. This isn't a new finding about a particular AI tool. For a short film made with automatic captions, I'd review the sounds separately from the words. Which sound explains a reaction, announces someone off-screen or changes what we know? Check that it reaches the captioned version when it happens. Describe what is audible; don't turn an unidentified noise into a named culprit the audience hasn't discovered yet. I'd ask a caption reviewer to watch the scene without audio, too. Can they follow why the character stops talking? A perfect transcript of the conversation wouldn't answer that.
Comments
If you're asking someone who relies on captions to review the film, pay them for that work and give them time before release. A friend agreeing to watch your short shouldn't discover they've been handed an evening of fixing the automatic captions. I'd send a version the editor has already checked, with an easy way to flag a moment by timestamp. Then the editor makes the corrections. The reviewer should get to talk about the film, too, not spend the whole conversation repairing access to it.
Give the reviewer the actual captioned player, not only a transcript to mark up. W3C's caption guidance says captions shouldn't obscure relevant information in the video; its media guide also warns that positioning support varies between players. In an imagined scene, a perfectly accurate caption could cover the hand slipping a key into a pocket. I'd check the version viewers will open, including on a phone, before calling the review finished. The editor should fix the placement and check it again there. Correct words in a caption file don't establish that the audience can still see the clue. This is a check suggested by existing guidance, not a fault I've reproduced in an AI captioning tool.