A few seconds of AI-generated video can now pass for a real recording, especially at phone size on a social feed. That clip of a storm, a celebrity or a "caught on camera" moment deserves the same questions you'd ask of a photo, plus a few that only apply to video.
1. Find the original upload
Reposted clips are cropped, re-encoded and stripped along the way. Look for the earliest copy you can find, and if you can, get the file itself rather than a screen recording of it.
2. Watch for what AI video still gets wrong
Play it at full size and slow it down if you can:
- Physics. Objects that pass through each other, liquids that pour oddly, people whose weight never quite lands.
- Continuity. Earrings, buttons, signs and background details that change or vanish between shots.
- Hands, teeth and text. Still the most common giveaways.
- Motion that is too smooth. Real handheld footage shakes, focuses and adjusts exposure.
3. Check what the file says about itself
A phone recording normally carries the device, the app, the creation date and sometimes the location. A generated clip starts without them, although a re-encoded real clip can lose them too, so treat missing details as a weak signal and contradictory ones as a strong one.
4. Run a check
Upload the video to snapWONDERS. Alongside the device and date details, the report reads any Content Credentials, checks the file's details for an AI declaration, and samples frames across the clip for a statistical estimate. It then gives the clip an AI-generation card:
- C, No AI-Generation Evidence Found is where most real clips land. Nothing pointed either way. It is not a finding that the clip is real.
- C, Inconclusive means something turned up that can't settle it alone.
- D means there is AI evidence, with a label saying what it was. F, AI-Generated (Certified) means signed Content Credentials from a trusted signer declare it AI-generated.
- A, Verified Camera Capture and B, Real Capture, Edited need signed Content Credentials that record a real camera capture.
A table below the card lists each check, what it found and how strongly it counted. The frame estimate is a weak reading. A video motion check is shown too, but it is not part of the score. How to read your report explains the card and the table in full.
A worked example: the same tram, two phones
We recorded trams at the same Melbourne stop a few moments apart, once on an Android phone and once on an iPhone (shared to us through Google Drive), and checked both clips.
We checked both again on 15 October 2026. Each reads C, No AI-Generation Evidence Found, with a statistical estimate of 15%, a weak reading, and none of the 8 sampled frames flagged. Here is the Android clip's AI section; the iPhone clip's reads the same:
When we first checked them in September 2026, the report called the same 15% "likely authentic". It no longer gives that estimate a verdict of its own. A clip with no AI evidence gets C, which says nothing pointed either way. It is not a finding that the clip is real.
Both real clips still grade Authenticity C ("moderate risk"), as they did in September. The reason is worth knowing:
The check samples 25 frames across the clip. These are steady shots seven or eight seconds long, so frames sampled a fraction of a second apart look almost identical, which is exactly what the check looks for when a clip has been padded with copied frames. The Android clip also flagged one sudden jump between two frames. Neither is a sign of AI. Short, steady footage gives these checks less to work with, which is why it pays to read the findings rather than stop at the letter.
See it as a before and after: two real tram clips
Reading the result honestly
- "No AI-Generation Evidence Found" is not "proven real". It means nothing pointed either way.
- Short, low-resolution clips give a check less to work with. Say so if that's what you have.
- Context still matters most. Who posted it first, when, and does anyone else have footage of the same event?
Related guides
More guides on
