Does X Auto-Generate Captions? How to Tell Auto From Uploaded
Yes, on some videos. We checked 23 public X videos: 10 had auto-generated caption tracks, 5 had uploaded ones and 8 had none. Here's how to tell them apart.

Yes, X auto-generates caption tracks on some videos. Not all of them. On 15 November we checked 23 public video posts from 7 to 14 November. 15 had a caption track and 8 had none. Of the 15 tracks, 10 were auto-generated and 5 were uploaded by the poster.
When you quote a clip, it helps to know which you have. An uploaded track came from the poster. An auto track is a machine's guess. What we found:
- X's video file names each track. Auto tracks are called
en (auto-generated). Uploaded ones are called "English" or carry an SRT filename. - Auto tracks have a second tell inside the file. All 10 auto tracks wrap each line in per-word timing tags. None of the 5 uploaded tracks had them, and 3 of the 5 started one hour in, at
01:00:00. - Coverage varied by account. All 10 WhiteHouse tracks were auto. All 5 NASA and NASAArtemis tracks were uploaded. Reuters had a track on 0 of 5 videos. Auto tracks come in Japanese too: 8 of 22 NHK News clips we checked had one.
- X's help pages disagree with each other. One lists auto captions for videos. Another lists them only for Spaces and Voice posts.
- A short track can still be complete. One auto track had 3 lines across a 28.7-second video. Two speech-to-text runs on our own machine found speech only in that stretch, about 4 seconds of it.
- ChatGPT cleaned an auto track well, with one slip. Logged out, on a 58-line track, it kept every timing and fixed the one misheard name. It also changed a word nobody asked it to change.
- Our tool can't tell you which kind it read. The X video transcriber reads both kinds and labels both "Original X captions".
What X's help pages say
X's page How to share and watch videos on X says: "Not all videos will have captions or subtitles available." The same page says: "Speech-to-text is generated using Microsoft Azure Cognitive Services."
After that it gets muddy. X's Accessibility at X page lists "Auto-caption support for videos, X Spaces, and voice posts". The Accessibility features of X page lists "Automatic captions (provided by Microsoft)" for X Spaces and Voice posts only. Videos aren't on that list.
Our sample backs the first page, at least for some accounts. We read these pages on 15 November. None of them shows a "last updated" date.
What we found in 23 videos
Every X video comes with a playlist file. The player reads it to find the video and any caption tracks. We opened that file for each post and read the caption line.
- 10 auto-generated, all from WhiteHouse. Each was named
en (auto-generated)and flaggedtwitter.auto-generated. - 5 uploaded, from NASA (4) and NASAArtemis (1). Three were named "English". The other two carried filenames:
jsc2026m000947-2026_Astronaut_Recruitment.en_US.srtandcaptions.en_US.srt. - 8 with no track: Reuters (5), SpaceX, OpenAI and YouTube.
Several of those 8 videos have clear talking in them, Reuters included. A missing track says nothing about whether anyone speaks.
Two things this sample can't tell you. First, why some accounts get auto tracks and others get none. X publishes no rule for it. Second, whether X's own player tells viewers a track is auto-generated. We didn't sign in to check. X's video file labels it; we can't say X shows you.
Two tells inside the caption file
Per-word timing tags. Every auto track we checked (10 of 10) wraps each line in a tag that starts <X-word-ms. It times each word separately. No uploaded track had it (0 of 5).
A one-hour offset. 3 of the 5 uploaded tracks started at 01:00:00, not at zero. On a NASA video, the first line sat at 01:00:00.375.
Our tool removes both tells. It strips the tags and corrects the offset, so the NASA line comes out at 0.375 seconds. The exported file won't show you either clue.
How to tell by reading the captions
You can still make a good guess from the words. These clues point to an auto track:
- Short lines of much the same length.
- No speaker names or sound tags like [music].
- Phrases split across lines. On the NASASpox video, "on" ends one line and "the moon" starts the next.
- Numbers spelled out, like "twelve" and "one million dollar".
Uploaded tracks often read as full sentences with punctuation. The NASA track opens: "Mr. President, welcome / to Kennedy Space Center."
These are clues. The label in X's file is the proof. The University of Colorado Boulder lists similar signs in Identifying Automated vs Human-Edited Captions.
What the X video transcriber shows you
Paste a public X post link into the X video transcriber and press Pick every word. You don't need an X login or a shinmuscat account. In our test, a video with a caption track came back in about 5 seconds.
The result card says Original X captions for both kinds. We checked both:
- Try an example loads a NASASpox video. Its track is auto-generated. The card read "Original X captions" and "58 segments · Captions extracted".
- A NASA video with an uploaded track also read "Original X captions", with 26 segments.
So if you need to know which kind you have, use the reading clues above. The label won't tell you.
One more thing. If you paste a post that quotes someone else's video, the transcript is credited to the account doing the quoting. Cite the original post yourself.
This guide uses videos that already have X captions.
Clean up an auto track with ChatGPT
An auto track is a draft. A chat AI can tidy it, as long as you check what it changed. We ran this in ChatGPT, logged out with no account. It may work for you without one too.
Step 1: Get the SRT
In the X video transcriber, open Download ▾ and pick SubRip subtitles .srt. Need a different format? Here's SRT vs VTT, and which file to download.
Step 2: Paste the prompt, the post text and the SRT
This is the prompt we used for the NASASpox video, word for word. The SRT went straight after it:
Below is the auto-generated caption file X made for a 146-second video posted by @NASASpox. The post text says: "From MAX POWER to Mars, it was a week to remember at @NASA! MAX POWER takes over @NASAKennedy. New Moon Base awards. X-70 Fractal unveiled. Astronaut recruitment opened. CHAPEA Mars simulation concludes. ESCAPADE sets off to Mars. Here's your NASA Minute!"
Clean this up so I can publish it as an accurate transcript:
1. Fix any words that look misheard (especially names, programme names and numbers), plus punctuation and capitalisation.
2. Keep it as an SRT with exactly the same cue numbers and timings. Do not merge, split or retime cues.
3. After the SRT, list every change you made as "cue number: old -> new", and mark any change you are not sure about with (unsure).
Put the post text in. It gave ChatGPT the right spelling of the name it fixed.
Step 3: Check what came back
58 lines went in and 58 came out. No timing or line number changed. X's own file had line 36 ending just after line 37 starts, and ChatGPT left that alone too.
It changed 41 lines. Most were punctuation and capitals: MAX POWER, Moon Base, Red Planet, PROMISE, ESCAPADE, Sun.
- One real fix. Line 37 said "NASA's Chappie habitat". ChatGPT changed it to "CHAPEA", which matches the post text. A speech-to-text run on our own machine also heard "Chapea".
- One change nobody asked for. Line 48 "towards" became "toward". Both of our speech-to-text runs hear "towards". That's an American style edit, and it changes what was said.
- Date and number style. "December 3rd" became "December 3". "twelve" became "12", and "one million dollar" became "$1 million".
- Its change list was honest. Every change it made was on the list. It marked only "CAPE Week" as unsure, and left it as it was.
ChatGPT also searched the web by itself and said it had checked some details against NASA's material. We haven't checked those claims, so we aren't repeating them.
Against our reference, X's auto track had one wrong word in 382. Our reference was two speech-to-text runs plus the post text. Nobody listened through by ear.
Step 4: If a track looks too short, ask before you fill it
We tried a sparse track as well: an auto track with 3 lines over a 28.7-second video. The prompt, word for word, with the SRT after it:
Below is the auto-generated caption file X made for a 28.7-second video. It only has 3 cues.
Clean this up so I can publish it as an accurate transcript:
1. Fix any words that look misheard, plus punctuation and capitalisation.
2. Keep it as an SRT with exactly the same cue numbers and timings. Do not merge, split or retime cues.
3. After the SRT, list every change you made as "cue number: old -> new", and mark any change you are not sure about with (unsure).
4. Tell me whether you think anything is missing from the transcript, and why.
ChatGPT kept the timings and made three small fixes. The biggest was "a follow-ups?" to "a follow-up?". Asked about gaps, it said: "Those gaps could contain silence, music, or non-verbal sounds, but they could also contain missing dialogue. Without hearing the original video, I cannot determine which." It added no dialogue.
That was the right call. Our two speech-to-text runs found speech only from about 10.4 to 15.3 seconds. The clip has around 4 seconds of talking, and the 3 lines covered all of it. Check the video before you fill gaps.
Next step
Paste a public X video link into the X video transcriber to check a video's captions free, then read the lines for the auto clues above. On TikTok instead? Here's what TikTok's auto captions can read, by language, and whether shinmuscat transcribes audio or reads captions. Auto captions mis-hear names, which trips up search. Here's how to find the exact moment someone said it in an X video.


