AI Captions vs Human Captions
accuracy is only half the problem

Three to five word chunks, one idea per card, the operative keyword emphasised in colour or weight, positioned above the platform UI. That pattern consistently outperforms full-sentence auto-captions in our own A/B tests, typically by 15 to 20 percent on average watch time.

It works because captions pace the viewer's reading rhythm. Full sentences dumped on screen invite skimming; short emphasised chunks pull the eye forward.

Burned-in captions serve sound-off viewing but do not satisfy screen readers or caption toggling. For accessibility compliance you also need an accurate SRT or VTT track, speaker identification on multi-speaker content, and non-speech audio cues where relevant.

We deliver both: styled burned-in captions for retention, and a clean caption file for accessibility and platform indexing.

Accuracy on clean, single-speaker English audio is now high enough that the raw transcript is a reasonable starting point. The failures cluster in predictable places: proper nouns, brand and product names, industry jargon, numbers and units, overlapping speech, and any accent under-represented in the training data.

For regulated industries the error categories matter more than the error rate. A transcript that renders a drug name, a legal term or a financial figure incorrectly is not a typo — it is a compliance exposure that sits burned into the frame of a public video.

Non-English and code-switched content degrades sharply. European Portuguese, Quebec French and heavily code-switched US Spanish all produce error rates well above the marketing figures quoted for standard English.

Even a perfect transcript makes poor captions if the timing is machine-derived. Automatic tools break lines on pauses or fixed character counts; a human breaks on meaning, so the punchline lands on its own line and the setup does not run over the cut.

Caption pacing also has to respect reading speed. Two lines of 32 characters held for at least 1.2 seconds is a workable floor; automatic output routinely flashes three lines for under a second during fast speech, which is functionally unreadable on a phone.

On short-form, the caption is often the hook. Where the first three words appear, how they are weighted, and whether the key word is emphasised are editorial decisions no transcription engine is making.

Machine transcription first, always — it removes the typing and gets timings roughly in place. Then a human correction pass focused on the known failure categories, then a re-timing and line-breaking pass against the edit, then a styling pass for emphasis.

That combination costs a fraction of manual transcription and produces captions that are accurate, readable and doing editorial work. Pure automation is defensible for internal or low-stakes content; it is not defensible on a page where a wrong number changes the meaning.