Auto captions are captions that software writes from the speech in a video, instead of a person typing them. A speech recognition model listens to the audio, turns it into words, gives each word a start and end time, and the platform shows those words as captions: either as a track viewers can switch on and off, or drawn into the picture. YouTube, TikTok, Instagram and most video editors offer them. They are a fast first draft, not a finished transcript.
Are they AI? Yes. The recognizer is a machine-learning model. They are not generative AI in the usual sense, because the job is to write down what was said, not to make new content. The catch is that modern recognizers predict text one piece at a time, so now and then they produce fluent words nobody spoke. That is why the platforms that offer auto captions also tell you to check them.
What auto captions are
"Auto captions", "automatic captions", "auto-generated captions" and "auto subtitles" all describe the same thing: text produced by automatic speech recognition (ASR) from a video's audio. YouTube's help page on automatic captioning, read on 19 October 2026, says it plainly: "YouTube can use speech recognition technology to automatically create captions for your videos". The same page lists 67 languages for automatic captions on long-form videos and Shorts, and English only for live streams.
Two distinctions matter more than the name.
- Captions vs subtitles. In US usage, captions are the speech written out in the same language, often with sound cues, and subtitles usually means a translation. In the UK, "subtitles" is commonly used for both. The captions vs subtitles guide has the details. Auto captions are almost always same-language text.
- Closed vs open. Closed captions are a separate track the viewer switches on. Open, or burned-in, captions are drawn into the video frame and always show. Most platform auto captions are closed; most styled captions you see on Shorts and Reels are open. See open vs closed captions.
Are auto captions generative AI?
They are AI. YouTube's page says its automatic captions "are generated by machine learning algorithms, so the quality of the captions may vary." Whether they are generative AI depends on what you mean.
In purpose, no. A caption model is judged on how closely its text matches what was said. It is not asked to write anything new, and a good one adds nothing.
In mechanism, partly. Modern recognizers are built like the language models behind chatbots. OpenAI's open-source Whisper model is described in its README as "a Transformer sequence-to-sequence model" whose output is "a sequence of tokens to be predicted by the decoder", and its December 2022 paper says it was trained on "680,000 hours of multilingual and multitask supervision". Predicting the next likely word is what makes these models fluent. It is also why they can fill a pause with words that were never said.
A 2024 study of Whisper, Careless Whisper, found that "roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences" in transcriptions made in mid-2023, and more often when speakers paused for longer. That is one model, at one point in time, on one set of recordings. It is not a figure for YouTube, TikTok or any captioning product.
For platform rules, YouTube treats captions as a minor use of AI. Its Disclosing use of GenAI content page, read on 19 October 2026, lists "Caption creation" among the examples creators do not need to disclose. If you post elsewhere, check that platform's own policy.
How auto captions work: audio to words to timings
Every auto caption system, from YouTube to a desktop editor, follows roughly the same steps.
- Audio out. The tool takes the audio track from the video. Nothing in the picture is read, so text already on screen is ignored.
- Language. Some tools detect the spoken language; others ask you to choose. A model set to the wrong language produces confident nonsense, so this setting matters.
- Speech to words. The recognizer converts short windows of audio into text. Whisper's README, for example, says it processes a file "with a sliding 30-second window".
- Words to timings. Each word or phrase gets a start and end time, so the text appears as it is spoken. Word-level timings are what make one-word-at-a-time captions and karaoke-style highlights possible.
- Timings to caption lines. The words are grouped into lines short enough to read, and each line stays on screen for a set time. Long lines, and lines that hang on screen after the speaker stops, are the usual signs of a poor split.
- Display. The captions are sent as a track the player can switch on, or rendered into the frame as burned-in captions.
Some platforms also filter the output. On YouTube, the "Don't show potentially inappropriate words" setting replaces flagged words in automatic captions with "[ __ ]" by default. You can turn it off in YouTube Studio under Settings, Channel, Advanced settings.
Where you see them
YouTube, TikTok, Instagram, LinkedIn and Facebook each build auto captions in differently, with different language lists, editing options and on/off controls. The platform auto captions comparison walks through each one.
What to check before you post
YouTube's advice, on the same help page, is to "Always review automatic captions and edit any parts that haven't been properly transcribed." It names mispronunciations, accents, dialects and background noise as common causes of errors. A five-minute check catches most problems. Play the video once with captions on and look for these:
- Names and brand words. A recognizer writes the closest common word for a name it has never heard. Check every name, product and place.
- Numbers. Fifteen and fifty, thirteen and thirty, prices and dates. One wrong digit changes the meaning.
- Words during silence or music. If the captions show "Thank you", "foreign" or "heat" over a quiet intro or a song, the model filled a gap. Why YouTube captions say "foreign" or "thank you" explains the pattern.
- Mixed languages. A sentence that switches language often comes out garbled in the language the model was not set to.
- Two people at once. Overlapping speech is usually merged or dropped.
- Bleeped words. "[ __ ]" on YouTube is a filter, not a recognition error.
- Line breaks and timing. Lines should be short, end where a phrase ends, and leave the screen when the speaker stops. If the captions lag behind the voice, see fixing subtitles that drift out of sync.
A vendor's "99% accurate" figure does not tell you how your clip will come out. How accurate AI subtitles are explains what that number measures and how to test your own clip in a few minutes.
Auto, edited and styled captions compared
"Auto captions" can end up as three different results, and a fourth kind skips the software entirely. The examples are as of 19 October 2026.
| Type | Who writes the words | Who checks them | How viewers see them | Example (19 Oct 2026) |
|---|---|---|---|---|
| Auto captions | Speech recognition | Nobody, unless the creator opens them | A track the viewer switches on | YouTube automatic captions, 67 languages |
| Edited captions | Speech recognition, then the creator | The creator | A track the viewer switches on | Automatic captions corrected in YouTube Studio |
| Styled burned-in captions | Speech recognition, then the creator | The creator | Always on, drawn into the frame | Animated captions on Shorts, Reels and TikToks |
| Human captions | A person | Usually a second person | A track, or burned in | Captioning services for TV, courses and legal work |
Auto captions are the cheapest and fastest. On a long YouTube video with clear audio they may be enough. Edited captions take a few minutes and fix the errors that matter. Styled burned-in captions are what short-form creators use, because the words show with the sound off and on platforms that do not caption for you, and they can highlight each word as it is spoken.
Where CapsAI fits
CapsAI is a tool for the third row. You upload an MP4 or MOV up to 10 minutes and choose the spoken language: English, Hinglish (Hindi-English written in Roman script), Hindi, or one of eight other Indian languages. There is no automatic detection on the caption uploader, so you pick it yourself. Speech recognition produces a draft with word-level timings. You correct any word and its timing in the browser, pick one of 37 caption styles, and export one MP4 with the captions burned in. It does not export SRT or VTT files, does not translate, and publishes no accuracy figure, so read the draft. CapsAI's English has been used most on Indian English; for other accents, run your own clip through the 3 free minutes and read the draft.
What to do next
If your platform's auto captions read well after one pass, use them and fix the names and numbers. If you want captions that are always on screen, checked word by word and styled, start with a short clip on the auto caption generator. There are 3 free minutes to start; packs are paid once, from $4 (₹49 in India, €4 in the euro area), and watermark-free export comes with every paid pack.
