How Accurate Are YouTube Auto Captions in 2026?

YouTube's auto captions have a bad reputation. You have probably seen the line that they are 60-70% accurate, or that one word in three is wrong. That number is old, and the system has improved since. Repeating it in 2026 is misleading.

The captions are mostly fine now. The problem moved somewhere else, and it gets less attention: the language they come in, and whether you can get the text out at all.

Short answer: Independent tests in 2025 and 2026 put YouTube's English auto captions around 85-95% accurate on clear, single-speaker audio, dropping to roughly 60-75% with heavy accents, music, or crosstalk. Google publishes no official figure. They now include basic punctuation and capitalization. The bigger problem is not the words. It is that YouTube captions a video in one language, and it picks which one.

How accurate are YouTube auto captions?

The 60-70% number comes from accessibility write-ups published years ago. It got repeated until it turned into received wisdom. It no longer describes clean English audio.

Independent tests from 2025 and 2026 land in a different place:

So the old number was not invented. It is just the worst case, not the average. One benchmark on fast-talking tech videos measured a median word error rate of 12.7%, which is about 87% accurate. YouTube's own help page says it keeps improving the speech recognition and warns that quality varies with accents, dialects, and background noise. Google has never published an accuracy number, so every figure above comes from third-party tests on small samples.

Two things follow from this.

First, modern auto captions include basic punctuation and capitalization. The old complaint about an unpunctuated run-on stream is out of date. What they still don't do is say who is talking. North Carolina's Department of IT puts it plainly: speakers "are not automatically identified or announced when they change." In an interview or a panel, the lines run together with no sign of who said what.

Second, the errors that survive are not the harmless ones. Meryl Evans is deaf and relies on captions, and she has spent years collecting the failures under the name she uses for them, autocraptions. Her sharpest example is a cooking video. The audio says "4 to 5 minutes. You should not preheat." The captions say "45 minutes. You should know to preheat the oven."

Every word in that caption is a real English word. Nothing looks broken. The instruction it gives is the opposite of the real one, and the cooking time is off by a factor of ten. This is what an accuracy percentage hides. At 90%, one word in ten is wrong, and you do not get to pick which one. Here it was "not". Evans's advice is to use auto captions as a first draft and edit them before publishing.

The real problem is the language

Here is the part that gets skipped. YouTube does not caption a video in whatever language you need. It generates captions in one language: the video's default language.

That has a few consequences that hit you the moment you are outside the video's home language.

The language has to be on the list. Automatic captions only exist for languages YouTube's speech recognition supports, which is about 60 of them. North Carolina's Department of IT lists the same limit from the other side: automatic captioning struggles with "accents, dialects, and languages with limited speech recognition support." If the video is spoken in something off the list, there are no auto captions at all. For live streams, automatic captions are English only.

The default language can be wrong. YouTube leans on the uploader's declared video language. When a channel's default says English and the video is actually spoken in Japanese, the wrong recognizer runs, and the captions come back as confident-looking nonsense. Creators hit this often enough that there are help-forum threads and how-to guides devoted to fixing it. Nobody publishes numbers on how often it happens, so treat it as a known failure rather than a statistic.

Here is what the menu actually offers on a video that has captions:

Off, Chinese (Simplified), Chinese (Traditional), English. That is the entire list. If you read none of those, the menu has nothing for you.

And YouTube acts on its guess. It now auto-dubs videos, adding an AI-translated audio track that plays by default. Digital Trends wrote up what that looks like in practice: a Marques Brownlee upload started playing in Japanese, apparently because there was Japanese text on the T-shirt in the thumbnail. The setting to turn dubbing off is in YouTube Studio. As the piece puts it, "Creators got a switch. Viewers did not."

The detection ran on thumbnail text, not on the audio. YouTube decides the language from signals that weak, and as a viewer you cannot override it.

Auto-translate only works if captions already exist

YouTube does have an auto-translate option in the captions menu, and it covers 100+ languages. It has one hard requirement that undoes it: there has to be a caption track already. Auto-translate translates existing captions. It does not listen to the video.

It also inherits everything upstream. If the recognizer misheard a product name, the translation faithfully carries the wrong name into your language. Every upstream error carries into the translation, and the translation can add its own on top.

So auto-translate only works on videos that already have captions in a major language. On uncaptioned videos it does nothing.

What if the video has no captions at all?

Plenty of videos have none. Auto captions do not appear on every upload. They can be off, unavailable for the spoken language, or still processing on something posted recently. Older videos and smaller channels often have nothing.

When there is no caption track, YouTube gives you nothing at all. The transcript panel is empty. Auto-translate has nothing to translate. There is no native way to turn that audio into text.

Why the transcript panel is not the answer

Most videos have a Show transcript button underneath, which looks like the answer. It isn't.

The transcript panel is built from the caption track. It is the same text in a side panel, so it carries the same recognizer errors and the same single language. Open the captions and the transcript on any video and you are reading one source twice.

The panel is also built for reading along while the video plays. There is no download button. Copying by hand drags the timestamps with it. Shorts have no transcript panel at all.

Captions vs transcript: what's the difference?

The two words get mixed up a lot. The difference matters here, because it explains why the panel disappoints people.

Captions only exist to help you follow the video while it plays. A transcript is text you can save and reuse anywhere. If you want to quote a video, study how it is written, or turn it into notes, you need that text version.

How to get a YouTube transcript you can actually use, free

1. Copy the video link. Hit Share under the video and copy the URL. Any public YouTube video works, including Shorts.
2. Paste it into GetAnyTranscript. Drop the link into the YouTube transcript generator and press Get transcript. No account required to try. It works whether or not the video has captions. When there is no caption track, speech recognition transcribes the spoken audio instead.
3. Read the result. You get clean, timestamped text in paragraphs. Copy it, or download it as plain text or Markdown.

4. Translate the whole thing. One click renders the transcript and its analyses into one of 10 target languages. This is the part YouTube can't do when a video has no captions to translate.

Same video, whichever language you pick, and the summary and analyses are translated along with it.

5. Use the extras. Summary, mindmap, viral moments, hook analysis, and titles all come from the transcript itself, with no extra input from you.

Short videos under 3 minutes are free with no sign-up, and you get 3 free transcriptions per month. If a run fails, it costs you nothing and you can retry.

The same process works for TikTok and Instagram Reels. We covered the TikTok version separately.

What this costs you

Accessibility. At 90%, auto captions still miss too much for deaf and hard-of-hearing viewers. Fixing that gap is the publisher's job. YouTube's own help page and North Carolina's IT department give the same advice: treat the automatic track as a draft, review it, and upload the corrected file.

Reach. A video captioned in one language is readable by the people who already speak it. Everyone else needs a translation that YouTube can only produce if the captions were there to begin with.

Trust. In a tutorial or a lecture, a mangled technical term sends viewers off searching for something that doesn't exist.