Updated: 2026-10-03
YouTube generates captions automatically for many videos, using speech recognition. They are free, available within hours of upload for most videos, and improved over the years. They are also the text underneath every transcript tool and AI summary that works on YouTube, SkimTube included. So how far can you trust them? This article does not give you one accuracy number, because there is none that holds across videos. It tells you what they handle well, what they get wrong, and how to check a transcript quickly before you rely on it.
No single accuracy figure
You will see figures online claiming that auto captions are 70, 80 or 95 percent accurate. We have not measured this ourselves, and we have not seen a published figure from YouTube that applies to all videos. Be careful with such numbers. Accuracy depends on the video, and the spread is wide: a clear single speaker in a quiet room is one thing, a panel discussion with crosstalk on a noisy street is another. A number averaged over both tells you little about the video in front of you.
What can be said is how it behaves, and that is more useful.
What auto captions usually handle well
- One clear speaker. A person talking directly to a camera, at moderate speed, with a decent microphone.
- Common vocabulary. Everyday words and phrases.
- Widely used languages. Recognition tends to be stronger for languages with lots of training material, and weaker for smaller ones.
- Clean audio. No music under the voice, no room echo.
If your video fits this description, you can expect a mostly readable transcript where the mistakes are small.
What they usually get wrong
Proper names. People, companies, places and products are the most common errors. A name can appear in three spellings in the same transcript, or turn into a similar-sounding common word. Brand names that look like normal words are especially risky.
Technical terms. Jargon, drug names, programming terms, scientific names. Recognition guesses the nearest common word.
Numbers. Digits, dates, prices and units are easily mixed up, and "fifteen" and "fifty" sound alike. If a figure matters, check it.
Accents and fast speech. Strong accents, mumbling and rapid talk lower accuracy a lot.
Several speakers. Overlapping voices become garbled lines. Auto captions typically do not mark who is speaking.
Mixed languages. A speaker who switches between two languages mid-sentence often breaks the recognition.
Punctuation and sentences. Auto captions are often a stream of lowercase text without proper sentence breaks. This does not change the words, but it makes reading harder and can hide where a sentence really ended.
How to see what kind of captions a video has
On a desktop browser, open the video settings (the gear icon) and choose the captions menu. A track labeled "auto-generated" was made by speech recognition. A track without that label was uploaded by the author or added by a captioning service. If there are several tracks, the transcript panel may let you choose between them.
An author-uploaded track is usually, though not always, better. It was either typed by the author or a captioning provider, or corrected after the automatic version. Some authors upload files that are just a rough draft, so verify before you trust them either.
A five-minute check before you rely on a transcript
You do not need to proofread the whole thing. Sample it.
- Pick three timestamps spread across the video, at the beginning, middle and end.
- Listen to 20 seconds at each while reading the matching lines.
- Count the errors. Zero or one per sample means the transcript is probably fine for notes. Several per sample means you should not quote from it without listening each time.
- Search for the names and numbers you plan to use and check each against the audio.
In SkimTube you can search the transcript and click a line to jump to the moment, which makes steps 2 and 4 quick. YouTube's own transcript panel does the same.
What this means for AI summaries
An AI summary reads the captions and writes from them. It does not listen to the video. So:
- If the captions say "cloud" where the speaker said "Claude", the summary may say "cloud".
- If a number is wrong in the text, the summary carries it forward.
- If a sentence is cut in the wrong place, the model may misread what the speaker meant.
This is why every tool that summarizes from captions should show you where in the video each point comes from. SkimTube puts a clickable timestamp on each key point for this reason. If you plan to quote a point, click it and listen. For how to write prompts that keep this check possible, see using a transcript with ChatGPT.
Using captions for quotes, subtitles and study
Quotes. Always listen to the line before you publish it. A wrong word in a quote is your error, not YouTube's.
Studying. A wrong term in your notes becomes a wrong term in your head. For technical subjects, check each key term against a textbook or the slides.
Language learning. Auto captions can teach wrong spelling. Prefer videos with author captions. There is more in using YouTube subtitles for language learning.
Searching. If a word you heard does not turn up, the captions may have written it differently. Try variants of the spelling and a few synonyms.
When there are no captions at all
Some videos have no captions: the author turned them off, the language is not supported, or the upload is too new for automatic ones to be ready. In that case there is no text to read and tools that rely on captions have nothing to work from. The guide on why the transcript is not showing goes through the checks. A speech-to-text service that listens to the audio is a different kind of tool and has its own accuracy limits.
Can you improve auto captions?
Authors can edit their captions in YouTube Studio, and many channels fix the worst errors. As a viewer you cannot change them. If a transcript matters to you, the practical improvement is to take the text, correct it by hand, and keep your corrected version. Save the file and fix the names and numbers once, so you do not hit the same errors again.
The short version
Treat auto captions as a good first draft. They are enough to find a moment, follow an argument or get a rough outline. They are not enough for quotes, numbers, names or technical terms without a check. Spend five minutes sampling any transcript before you build on it, and use the timestamp to listen to anything that matters.