motion field guide
How to Spot AI-Generated Videos
Learn how to inspect motion, physics, continuity, and fine detail in AI-generated videos, with annotated real and synthetic examples.
AI video often looks convincing one frame at a time. The illusion tends to break when you follow an object, hand, reflection, or background detail through several seconds of motion.
Use these examples as comparisons, not as a universal detector. Generators improve quickly, and unusual real footage can look just as implausible as synthetic video.
Inspection checklist
Follow one object
Choose a hand, face, wheel, shadow, or background sign and track it across several frames. AI video may quietly change its shape or identity.
Check physical reactions
Look at weight, momentum, splashes, fabric, hair, and collisions. Motion can be smooth while the underlying physics makes no sense.
Watch the edges
Hands crossing objects, hair against a background, and overlapping limbs are common places for temporary melting or impossible occlusion.
Listen to your doubt—then verify
A cinematic or uncanny feeling is only a reason to inspect more closely. Compression, stabilization, and unusual lenses can also make real footage look synthetic.
from the game
Annotated examples
Fan in the stands walks onto the pitch
Watch the match clock in the top-left corner: over 15 seconds it jumps around between 34:05 and 34:09 instead of counting up. Then the story falls apart: the woman eating a burger in the stands walks onto the pitch mid-match, kicks the ball among the players, and waves at the camera. The creator (@Ciri_ai) disclosed it was made with Seedance 2.0 and GPT Image 2, but most of the 2.5M viewers never saw that.
I don't think this is AI, but everyone says it is
Real. A farmer (@Wanderingmahesh) opens irrigation channels with a hoe and water floods down each furrow. Small things stay consistent for all 23 seconds: a scrap of paper in the soil behaves like paper, and bits of straw carried by the water stay intact. AI video tends to morph or drop small objects within a few frames. At the end the cameraman's shadow falls across the field. The comments were sure it was AI, and they were wrong.
Happyhorse — "even AI detectors can't tell"
It's framed as if filmed from the street through a restaurant window, but in the last shots the glass is gone: no reflections, no frame, just the couple filling the screen. Then both turn and stare into the lens at the same moment. It was posted by an AI consultant promoting Alibaba's HappyHorse video model ("this looks 100% real, even AI detectors can't tell"), and it later spread on X as supposed footage tied to the 2026 JPMorgan harassment lawsuit. It's AI.
Wife thinks this is AI — rock trajectory discrepancy
Real. The rock hits the sea closer to the plane than you'd expect because it keeps the plane's forward speed as it falls. It isn't left behind; it travels along underneath. The 19-second clip is one continuous shot: the seat, cables and scratched ramp stay identical as the camera swings around, and the video predates AI video tools. "The physics looks wrong" sometimes just means the physics is counter-intuitive.
AI short film — "production costs from millions to hundreds"
This 67-second retelling of the Hachiko story is stitched together from short shots, most only 3 to 7 seconds long. AI video models generate brief clips, so AI films rarely hold a long continuous action. After "10 years later", the owner doesn't look older; he looks like a different, much older man. It was posted by an AI consultant claiming "AI has reduced film production costs from millions to hundreds."
Circassian dance — "the famous Russian dance" is actually Circassian
Real. The dancers seem to glide across the floor without taking steps. That floating look is the signature of traditional Circassian women's dance, done with tiny steps hidden under floor-length skirts. It's a 75-second rehearsal in an ordinary studio: worn floorboards with red tape marks, New Year baubles hanging from the ceiling, a Christmas tree in the corner, and a mirror that reflects the room correctly. Dozens of faces stay consistent the whole time.
"Kiss" — horse and woman
Watch the horse's head: as it leans in, the muzzle stretches and bends like rubber, and the white marking on its face changes shape. The woman's cape is made of hundreds of identical petal shapes. AI creative director @gizakdag posted it with a single word, "Kiss." Beautiful, strange and uncaptioned is a common formula for AI clips made to be shared.
Feelgood Instagram channel — real or AI?
Real. The dog's odd behaviour is trained medical-alert work: jumping up at her at the dishwasher, getting her down onto the floor, going to the fridge, then lying across her. This is @serviceaussiebailey, a service dog that detects seizures and POTS episodes. It's almost two minutes from a fixed camera with no cuts, and the kitchen never shifts. A feelgood page reposted it without the medical context, which made it look uncanny.
Guy AI-inserts himself into Game of Thrones
He's in a hoodie, shirt and tie in Westeros, handing Jon Snow a pink balloon sword. The joke is the tell. Then look at the familiar actors: their faces are close to the originals but slightly off. @Rixhabh__ used AI to put himself into Game of Thrones to "fix everything": comforting Ned Stark at his execution and crashing the Red Wedding.
Can ducklings move like this? Seen on Instagram
Real. A mother duck leads her ducklings down a steep concrete dam spillway. It looks wrong because we rarely see it, but ducklings are light and fluffy and grip rough concrete well, so they tumble and slide without getting hurt. The shaky handheld camera pans and zooms freely through one continuous 32-second clip. Motion blur on small, fast bodies is normal camera physics, not an AI artifact.