Your Auto-Captions Can Invent Words: How to Proof an AI Transcript

The Caption Said Something Nobody Said

Automatic captions are usually right, and that is the problem. A creator checks the first two clips from an episode, sees clean text, and stops checking. Then a clip ships with a word the guest never used, burned into the frame where no viewer can switch it off and no comment can correct it.

Speech recognition fails in two different ways, and they need different checks. The familiar failure is mishearing: a name, a number, a brand spelled the way it sounds. The less familiar one is invention. OpenAI's own model card for Whisper says that "the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination)" (Whisper model card).

How Often a Transcript Invents Text

A research team measured it. Evaluating Whisper, the authors of "Careless Whisper" found that "roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences which did not exist in any form in the underlying audio," and that "38% of hallucinations include explicit harms such as perpetuating violence, making up inaccurate associations, or implying false authority" (Koenecke et al., ACM FAccT 2024).

Three things about that finding matter for a clip workflow:

• The unit is a transcription, not a word. One in a hundred sounds rare until you multiply it by every segment of every episode in a season.
• Gaps are where it happens. The paper found that hallucinations "disproportionately occur for individuals who speak with longer shares of non-vocal durations" (Koenecke et al.). The study was about speakers with aphasia, but the practical reading for clip makers is plain: look hardest at pauses, dead air and music beds.
• It is a measurement of 2023-era Whisper. The paper describes the model it tested as state of the art "as of 2023." Models improve. The habit of checking should not depend on which version you happen to be running.

Mishearing Is the Everyday Failure

Invention is the dramatic case. Mishearing is the common one, and it is not spread evenly. The same model card says Whisper models "perform unevenly across languages" and "exhibit disparate performance on different accents and dialects of particular languages" (Whisper model card). It also warns that the architecture "makes it prone to generating repetitive texts," which shows up in captions as one phrase looping three times.

YouTube gives its creators the same instruction for its own automatic captions: they "might misrepresent the spoken content due to mispronunciations, accents, dialects, or background noise. You should always review automatic captions and edit any parts that haven't been properly transcribed" (YouTube Help).

The W3C's accessibility guidance has the clearest example of why a small error is not a small problem: "missing just one word such as 'not' can make the captions contradict the actual audio content" (W3C Web Accessibility Initiative). A guest who said "you should not do this" becomes a clip telling a hundred thousand strangers to do it.

Why Burned-In Captions Raise the Stakes

Platform captions can be edited after publishing. Captions burned into the frame of a short clip cannot. Fixing one means re-exporting and uploading a new post, and a new post starts from zero: the views, comments and shares stay on the broken one. The only cheap moment to catch an error is before export.

A Five-Minute Proofing Pass for Every Clip

This is the routine we recommend. It is short enough to run on every clip, which is the point: a check you skip on busy days is not a check.

1. Watch once at full speed with the sound on. Read the captions, do not just glance at them. Your ear catches what your eye accepts.
2. Check every proper noun and every number. Guest names, company names, product names, prices, dates, percentages. These are the words a model is most likely to spell by sound, and the ones a viewer is most likely to repeat.
3. Check every negation. Not, never, no, can't, won't, don't. One missing word flips the meaning.
4. Look hard at the edges and the gaps. The first and last second of the clip, every pause, every place where music or crosstalk sits under speech.
5. Watch for loops. The same phrase twice in a row is usually the model, not the speaker.
6. Keep a glossary per show. Names and terms that recur every episode should be fixed once, at the source, so every clip cut from that transcript inherits the fix.

If the clip makes a claim of fact, check the claim as well as the caption. A caption that faithfully reproduces a guest's wrong number is accurate and still a problem.

Where ClipForge Fits

ClipForge's default transcription model is OpenAI's whisper-1, and its clip detection reads that transcript to choose and score moments. That means a transcript error can travel to two places: the captions on the clip, and the text the detection step reasons about. The proofing pass above is the right habit with ClipForge as with any tool built on speech recognition. ClipForge has not published a caption accuracy figure, and this post does not invent one.

The User Guide shows where a project keeps its transcript, its clips and their exports, and pricing lists which caption styles each plan includes.

Keep Reading

• Why Auto Captions Are Essential for Short-Form Video in 2026
• Captions That Count: What WCAG and the ADA Ask of Short-Form Clips
• How AI Clip Detection Works: The Technology Behind Viral Moments

, Rocky