Security
How to tell if a video is AI-generated
Video was the last format people still trusted by default. A photo could be faked, a voice could be cloned, but footage of an event felt like evidence. That gap closed faster than almost anyone expected, and the advice that's circulating — watch the blinking, check the hands — is aging the same way the AI-photo checklist did. Here's what actually holds up.
Why the visual tells keep expiring
Every list of giveaways describes the weaknesses of one generation of models. Unnatural blinking, warping hands, hair that behaves like liquid, text on signs dissolving into symbols, physics that's subtly wrong — all of these were real, and all of them are being fixed, because they're exactly what the people building these systems are measuring themselves against.
So a checklist of tells has a shelf life of months, and worse, it fails in the dangerous direction: you check the six things you learned, find nothing, and conclude the video is real. An absence of known flaws is not evidence of authenticity — it might just mean a newer model. This is the identical trap covered in how to tell if a photo is AI-generated, and the fix is the same. Stop interrogating the pixels. Interrogate the provenance.
The remaining tells are still worth a glance — text and small background details are genuinely still hard, faces of people you know well can read subtly wrong, and audio that doesn't quite sit with the mouth movement is a real signal. Treat them as a reason to look further, never as a verdict.
Check one: where did this come from?
The most reliable question isn't about the video, it's about its history. Real footage of a real event has a trail.
- Who posted it first? Not who you saw it from — trace it back. An account created recently, or with no history in the topic, posting dramatic footage of a major event is the single strongest warning sign there is.
- Is anyone else covering it? A genuine newsworthy event has multiple angles, multiple accounts, and news organisations following within hours. A single clip, existing nowhere else, showing something that would obviously be news, is almost always synthetic or miscaptioned.
- Reverse image search a frame. Screenshot a distinctive moment and run it through Google Lens or TinEye. This catches the most common deception by far, which isn't AI generation at all — it's real footage from a different event or a different year, recaptioned.
Check two: content credentials
The same C2PA standard covered in the photo article applies to video, and adoption is growing. Many AI video tools attach a record of how a clip was made, and some cameras and editing tools attach one for real footage. You can inspect it with the free verify tool at contentcredentials.org.
Google's SynthID watermark works similarly for content from its own tools, embedded in the pixels rather than the file metadata, so it survives re-encoding and cropping better than a metadata tag.
The honest caveat is exactly as before, and it matters more for video: a found credential is strong evidence; a missing one tells you nothing. Most video on the internet has no credential, uploading to a social platform frequently strips metadata, and plenty of generation tools don't attach anything. Use it as confirmation, never as a test.
Check three: the question that beats all the others
Before any of the above — what is this video trying to make me do?
Almost nobody fakes a video for its own sake. It's made to move you: to donate, to be furious at a group of people, to buy something a celebrity appears to endorse, to believe a politician said a thing days before an election, to trust an investment because a familiar face is recommending it.
That last category is worth naming specifically, because it's now routine: a cloned face and voice of a well-known person promoting a trading platform or a giveaway. If a famous person appears to be offering you money, it is a scam, and no amount of frame-by-frame analysis is a better use of your time than that single rule.
The urgency is the attack — the same structure as an AI voice clone phone call and an AI-written scam email. The medium changed; the shape of the manipulation didn't.
The 60-second check
1. Is this video pushing me toward money, outrage, or a
fast decision? If not, don't bother.
2. Who posted it FIRST, and do they have history here?
3. Is anyone else showing this event, from another angle?
4. Reverse-image-search one frame. Old footage, new caption?
5. If it still matters: check contentcredentials.org.
6. No result either way is not proof. Verify through a
source the poster doesn't control before you act.Keep your head:
Watching more closely is no longer the safeguard, and a checklist of visual tells expires every few months. Trace where the footage came from, check whether anyone else has it, and ask what it wants from you. If a familiar face is offering you money, that's the whole analysis.
Get one of these a week. Our free newsletter sends one genuinely useful AI habit and one judgment check every week — no hype, four-minute read. Subscribe on the home page.
Related: How to tell if a photo is AI-generated and How to tell if that phone call is an AI voice clone.