Things change behind other things
A person walks behind a post and comes out in a different shirt; a car turns into a different car; a face changes after a head turn. Video models keep only a loose memory of what was there a second ago.
Guide
To tell if a video is AI generated, slow it down and step through it. AI clips look convincing at full speed and fall apart frame by frame: things change when they pass behind other things, text shifts, feet slide, and shadows move on their own. Check where the clip came from and whether it carries a provenance label, then let the frames do the talking.
Try it: six frames from a made-up AI clip. Drag the slider and find what changes.
Frame 1: a person walks past a bakery holding a red cup.
A person walks behind a post and comes out in a different shirt; a car turns into a different car; a face changes after a head turn. Video models keep only a loose memory of what was there a second ago.
A cup leaves a hand without being put down, a dog gains a leash, an extra chair arrives. Count the objects at the start and at the end.
Signs, T-shirt prints, number plates and screens whose letters drift or rearrange between frames.
Feet that slide over the ground, hair that moves without wind, liquid that pours too slowly, a ball that bounces without losing height, cloth that passes through a body.
Still frames may look fine; hands that grab, type or wave can merge fingers, pass through objects or change size.
Smooth, poreless skin, teeth that blur into one shape, blinks at odd moments, and mouths that move a little apart from the words.
People in the back who repeat the same motion, walk without getting anywhere, or melt into the crowd.
A perfect drone glide through a doorway, an endless slow push-in, or “phone footage” that is too smooth and too well framed.
Voices with no breath, room sound that never changes when the camera moves, footsteps that miss the feet.
Most AI tools make clips of about 5 to 15 seconds, and a few reach 30. A long video made of many shots of the same length, none longer than that, is worth a closer look.
Tick only what you actually saw. The meter weighs the signs; it is a guide, not a verdict.
Nothing ticked yet. Start with the source.
A clip that puts words in a real person’s mouth needs more care than a generated landscape. Look at the mouth at a quarter speed: lips, teeth and tongue often lag behind or blur during fast words. Watch where the face meets the hair and the neck, and look for lighting on the face that does not match the room.
Then step away from the pixels. Did the person’s own channels post it? Did any news outlet that was in the room report it? A statement that matters will usually leave a record somewhere other than one viral post.
Wurt.app is a browser app that turns a prompt or a picture into a video, with sound on several models and clips up to 30 seconds. Make a few clips, step through them, and you will learn where to look. Free to start, then pay as you go, with no subscription.
Open Wurt.app
Find where the clip was first posted, check it for a provenance label, then watch it slowly. Step through it frame by frame and look for things that vanish or change behind other objects, text that shifts, feet that slide, and shadows that move on their own. Several tells together are strong evidence.
Things that change when they pass behind something. A person walks behind a post and comes out in a different shirt, a cup disappears from a hand, a car changes model in the background. Real cameras never do that; AI clips still often do.
Most AI video tools make clips of about 5 to 15 seconds, and a few reach 30 seconds. A longer video built from many short shots of the same length, with no shot longer than that, is worth a closer look.
Yes. Many current models generate sound and speech together with the picture, so sound alone no longer proves a clip is real. Listen for voices with no breath, room sound that never changes, and footsteps that miss the feet.
Partly. The Content Credentials verify site reads signed history attached to some files, and Google’s SynthID check in the Gemini app looks for the invisible signal in clips made with Google’s AI tools. Neither can clear a clip that carries no label, so watching closely still matters.