How to Spot AI-Generated Footage (Without Fooling Yourself)

Generative video has improved quickly enough that the old advice — count the fingers, look for melting text — now fails often and fails in both directions. The practical approach is a layer of checks: where the clip came from, what the motion does, what the details say, and what the file’s provenance claims. None is decisive; together they tell you how much confidence the evidence can support.

Start with the source, not the pixels

Before studying a frame, find out who posted the clip and when. Search for the earliest upload, then read the account’s history: does it normally post this kind of footage, and is it the original source or a repost? Look for the surrounding article or post, because a clip presented as breaking news and a clip presented as a demo carry different burdens of proof. Check whether other outlets have independently verified it, and whether the file has been re-uploaded many times. Re-encoding and reposting destroy the small visual cues that people rely on.

Ask what the poster gains: an account chasing engagement, a company selling a tool and a newsroom with a corrections policy are not equally reliable narrators, whatever the pixels show.

Check motion and physics

Watch the clip at normal speed and then frame by frame. Generated footage often struggles with contact and weight: feet that slide, hands that pass through objects, objects that change mass when picked up. Liquids, smoke, cloth and hair are common failure points, as are crowds and background characters, which may drift, merge or repeat. Look at reflections and shadows for consistency, and at the camera itself: a real camera has a physical shutter, a lens and a stabiliser, and its movement usually has a continuous cause.

Real footage also produces impossible-looking moments — a rolling shutter, heavy compression, slow motion at a low frame rate. A single physics error is a reason to look closer, not a verdict.

Look at detail under pressure

Text, hands, teeth, ears, jewellery, repeating patterns and the edges where hair meets background are all places where generative models have historically slipped. Text may be almost legible rather than actually readable. Fingers may bend oddly or merge. Background faces may be blurred in a way that hides structure, or repeated with small differences. Edges may have a soft halo that does not match the depth of field.

The caution is that these tells date quickly. Current tools handle hands and short text far better than the tools of a few years ago, and a careful operator can fix many artifacts in post. Use detail checks to generate suspicion, then verify elsewhere.

Check the light and the optics

A convincing fake has to reproduce the behaviour of a lens: depth of field that matches the focal length, motion blur that matches the shutter, sensor noise that stays consistent across the frame, and exposure that responds sensibly when the camera moves between light and shadow. Generated clips sometimes show a sharpness that is uniform across the image, or grain that changes from shot to shot, or lighting directions that disagree between a face and its background.

Real footage can look wrong for the same reasons: heavy denoising, upscaling, filters and regrading can flatten exactly the cues you are checking. A clean image is not evidence of authenticity.

Listen to the audio

For a talking-head clip, the voice is often the fastest route to doubt. Listen for breath, mouth noise, room tone and the small overlaps of natural speech. Synthetic voices can sound clean in isolation but odd against a room: no breath before a long sentence, ambience that does not change when the speaker turns, or lip movement that leads or lags the words. Voice cloning is a separate problem from video generation, and either can be faked independently.

Read the provenance

Some cameras, phones and editing tools can attach signed provenance metadata. C2PA is an open technical standard for recording where content came from and how it was edited; Content Credentials is the name used for that metadata in consumer tools, and verifying services can display the signer and the edit history. For footage, this is the closest thing to a chain of custody: it tells you what a device or application claims, and who signed the claim.

It is not a truth serum. Metadata can be removed, platforms frequently strip it, many cameras do not sign anything, and a signed clip can still be mislabelled. Absence of credentials proves nothing. Presence is useful evidence, especially when it names a known camera or publisher and shows a plausible edit history.

What the tells cannot prove

Every signal above is probabilistic, and the honest conclusion is usually a confidence level rather than a verdict. Real footage can display artifacts, especially after compression, denoising, upscaling or repeated reposting. Generated footage can be immaculate, especially in a short clip with little motion. Detector tools exist, but they are trained against particular generators and produce scores rather than proof; a new model or a light regrade can move a clip across a detector’s boundary. And there is a further problem: once people know fakes exist, real footage gets dismissed as fake, a dynamic sometimes called the liar’s dividend.

The workflow below puts source checks before pixel checks. The question is rarely “is this synthetic in some absolute sense?” It is “what should I believe, and what should I say I believe?”

A ten-minute check

1. Find the earliest version. Search the clip’s distinctive phrase or image, and note the oldest account that posted it.
2. Read the account and the surrounding post. Identify the incentive and whether the poster claims to be the source.
3. Check independent coverage. Look for a newsroom or specialist verifier that has examined the clip.
4. Watch once for story, once for physics. Note specific moments that look wrong, with timestamps.
5. Look for provenance. Check for a Content Credentials pin or metadata viewer, and read the signer.
6. Say what you know. If nothing is confirmed, describe the clip as unverified rather than as fake or real.

If you are still unsure

Do not present the clip as fact, and do not present it as a proven fake either. Ask the person who posted it where it came from, and keep a record of the answer. Share the earliest source rather than a repost. If you have already shared something that turned out to be synthetic, correct it in the same place and explain the tell — quiet deletion teaches nobody.

The skills are ordinary media literacy applied to faster tools. For how these systems generate footage in the first place, see how AI video models work. For the vocabulary, the AI cinema glossary covers the terms in plain language. And for an example of AI-made work that is open about what it is, browse the guide to AI-generated short films.