A clip starts spreading on X. Maybe it’s footage from a protest, a CEO saying something careless on stage, or a video that looks a little too perfect to be real. By the time someone decides it’s worth checking, there’s a decent chance the post has been deleted, the account suspended, or the video edited and reposted somewhere else with a new caption.
For anyone whose job involves verifying what they see online, that’s the real problem. You can’t transcribe a video that no longer exists, and you can’t run a deepfake check on a screenshot of a deleted post. Every AI tool in a modern verification workflow, from speech-to-text models to frame-by-frame analysis, needs the original file to work with.
Why the file comes first
X doesn’t make saving a video simple for most people. Since 2023, it has offered a download button, but only to paying subscribers, and only on videos whose creators allowed downloads. That leaves out every older post, any creator who switched the option off, and anyone without a subscription.
So researchers and reporters often turn to third-party downloaders. Tools such as RememberThisX take the link of a public post and return the video as a standard MP4, without an account. That convenience comes with obligations, covered further down.
The file matters because of what you lose without it. A screen recording re-compresses the video and drops quality. A screenshot keeps one frame and none of the audio. A link is only good until the post is deleted. The original file keeps the most evidence.
A workflow that holds up later
Open-source investigators worked much of this out years before AI made it urgent. The Berkeley Protocol on Digital Open Source Investigations, published by UC Berkeley’s Human Rights Center with the UN Human Rights Office, sets standards for identifying, collecting, preserving, verifying and analysing online material. It was written with war-crimes investigations in mind, but the logic carries over to any newsroom.
Here’s how that translates to a single X video:
| Step | What to do | Why it matters |
| Capture | Save the video file early, at the highest resolution available | Viral posts are often the first to be deleted |
| Document | Record the post URL, account handle, post time, capture time, and your method | Without provenance, the file is a claim, not evidence |
| Fingerprint | Generate a SHA-256 hash of the file | Proves later that the file hasn’t been altered since capture |
| Transcribe | Run the audio through a speech-to-text model | Makes the clip searchable and quotable |
| Analyze | Pull key frames, reverse-search them, check for manipulation | Finds earlier versions and signs of editing |
| Verify | Confirm claims against primary sources | AI summaries are leads, not findings |
The hash step sounds technical but takes seconds. On a Mac, shasum -a 256 filename.mp4 in Terminal does it. On Windows, it’s certutil -hashfile filename.mp4 SHA256. Write the result down next to the URL and capture time. If anyone later questions whether you edited the clip, the hash settles it.
Where AI helps, and where it trips
Transcription is the easy win. Modern speech-to-text models handle short, noisy clips far better than they did a few years ago, and a transcript turns a 40-second video into something you can search, quote, and translate. One caution is well documented: some models occasionally invent words or whole phrases that were never spoken, particularly during silences or poor audio. The Associated Press reported this problem with OpenAI’s Whisper in 2024. Before you quote anything, listen to that part of the clip yourself.
Summaries come next. Language models are good at condensing a pile of clips and threads into a short overview. They’re also good at sounding confident about things they got wrong, and they can be led astray by a single planted source. One journalist showed how a fake web page fooled both ChatGPT and Gemini within a day. Use the summary to decide what to check, then go and check it.
Manipulation checks are where having the original file matters most. Pull a handful of key frames and run reverse image searches. That often turns up the same footage posted years earlier, or in another country, under a different story. Check whether the audio lines up with the speaker’s lips, and whether shadows, text, and reflections behave as they should.
Be careful with “deepfake detectors,” though. Synthetic video has improved quickly, and tools like Google’s Gemini avatars can now simulate a person rather than just paste their face onto someone else’s. Detection tools struggle to keep up and produce both false alarms and misses. Treat any detector score as one clue among several, never as a verdict.
Content credentials, such as the C2PA standard some cameras and AI tools now attach to files, can help when they’re present. Social platforms usually re-encode uploads and strip much of the original metadata, though, so a downloaded X video often won’t carry them. Missing credentials prove nothing either way.
Quality has a ceiling
AI analysis is only as good as its input. Heavy compression blurs the small artifacts that give away an edit, and muddy audio drags down transcription accuracy. So save the best version X actually kept. Some downloaders, RememberThisX among them, offer up to 1080p when the original upload was that sharp.
No tool can recover detail X never stored. If someone uploaded a 480p clip, 480p is the best anyone will get. A downloader promising more than the original is upscaling, which is the last thing you want in a verification workflow, since it adds detail that was never there.
The legal and ethical side
Saving a public video for reference, research or reporting is a very different act from re-uploading it as your own. A few principles keep the line clear:
- Stick to public posts. Private or protected accounts are off-limits, and a tool claiming to get around that is a red flag.
- Credit the creator whenever you quote or describe a clip.
- Don’t republish or monetize footage you don’t have rights to.
- Honor takedown requests from rights holders.
Copyright is only half of it. X’s own terms say users may access the service only through its published interfaces, and they ban crawling or scraping in any form without X’s prior written consent. Whether those terms are enforceable against one-off saves of public videos is a contested legal question, and courts haven’t settled it. Newsrooms and research teams should have their own policy rather than assuming it’s fine because everyone does it. Fair-use rules also vary by country, so what’s allowed for reporting in one place may not be in another.
For a step-by-step walkthrough on iPhone, Android, Mac and PC, plus more on the copyright questions, the complete guide to downloading X videos covers each platform.
Questions people ask
Q. Can’t I just bookmark the post?
A bookmark points to the post. If the post is deleted, the bookmark points to nothing. Only a saved file, with your notes and hash, survives.
Q. Is a screen recording good enough?
For a casual reference, maybe. For verification, no. It re-compresses the video, can change frame timing, and makes the file harder to authenticate later.
Q. Can AI tell me if a video is fake?
It can flag things worth examining. It can’t give you a reliable yes or no on its own. Combine detector output with reverse searches, source checks, and the original context of the post.
Final Thoughts
As AI makes convincing fake media cheaper to produce, the boring step of saving the original file properly, documenting it and fingerprinting it becomes one of the most valuable skills in verification. Everything after that depends on it.
About the author: [Author name] is the founder of RememberThisX, a free web tool for saving public X videos. This is a guest contribution; AIInsightsNews does not endorse specific tools.
Related: AI Video Analytics: Your Video Got Views, But Did It Reach the Right People?
