Source Catcher

How to tell if a video is AI generated

What actually gives it away, what does not, and why "not certain" is often the only honest answer.

Before you start

This is the least reliable check on a video, and it should be your last. Most video misinformation is not generated at all — it is real footage with a false caption, or old footage relabelled as new. Those are far easier to disprove and far more common.

So ask what the clip was cut from and how old it is first. Most clips are settled before you ever need this page.

What helpsSigns worth actually looking for

Generated video has got dramatically better at faces and dramatically less bad at hands. What it still struggles with is consistency over time and things with rules.

Text in frameThe most reliable tell there is. Signs, labels, number plates, shop fronts, writing on clothing. Generated text tends to be almost-words, or to change between frames.

Things that should persistWatch one detail across the whole clip — a logo, a scar, a pattern on a shirt, the number of buttons. Drift over time is very hard to fake away.

Where hair meets backgroundEdges are still hard. Look at hair against a busy background, and at where fingers cross in front of objects.

Physics with rulesWater, smoke, cloth and crowds all behave in ways models approximate. Watch whether a reflection matches what is in front of it.

Teeth and jewelleryCounts change. Earrings appear in one shot and not the next.

Gait and hands at restPeople walking away, or hands doing nothing in particular, are often where it falls apart.

Suspiciously clean audioSpeech with no room tone, no breath and no background is worth a second listen.

What does not helpThe mistake that gets real footage called fake

This is the more common error, and the more damaging one. Calling genuine footage AI is as harmful as missing a fake, and it happens far more often than people expect.

"It looks off." Compression makes real footage look wrong constantly. Low light, heavy re-encoding from repeated reposting, aggressive phone processing and beauty filters all produce exactly the smeared, waxy look people read as AI.
Smooth skin. That is a filter, and half the internet uses one.
Odd frame rates. Motion smoothing and slow-motion do this.
"Too perfect." Professional lighting exists.
Detector scores. See below.

Be sceptical of AI detectors

Tools that return a confident percentage are the least trustworthy thing in this whole area. They are trained on particular generators, they degrade badly on compressed social video, and they produce false positives on genuine footage with no warning. A number like "94% AI" reads as evidence and usually is not.

Treat any detector output as one weak opinion, and never as the answer.

What would settle itContent credentials, and why they are usually gone

There is a real technical answer to this problem: C2PA content credentials, a signed record of how a file was made and edited, attached to the file itself. Cameras and generators increasingly write it.

The catch is that platforms strip it on upload. By the time a clip reaches you through TikTok, Instagram, X or a WhatsApp forward, the credential is gone. So its absence proves nothing at all — the overwhelming majority of genuine video you see has none. When it survives, it is strong evidence. When it does not, you have learned nothing.

How to say what you found

Unaided, most people score between about 55% and 75% at spotting generated video — barely better than a coin toss. That should make everyone more careful about announcing verdicts.

Good: "The sign on the shop changes spelling twice. That points to generated."
Good: "I can't tell. The compression is heavy and nothing in frame is decisive."
Bad: "It's obviously AI, look at it."

Citing what you actually saw lets someone check your reasoning. "It looks fake" cannot be checked by anyone, including you.

Get an honest read on a clip — free, no sign-up

Source Catcher watches the video and gives its read on whether it looks generated, citing what it actually saw rather than returning a score. It says "not certain" when it is not certain. It also does the checks that matter more: what the clip was cut from, whether anything independent confirms that, when it first appeared online, and whether the caption honestly describes the footage.