Why AI video glitches
Lesson 2 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
Five glitches that give AI away
Mira made her first clip and it looks pretty. But the comments instantly say “AI slop”. Viewers catch the same mistakes every time: extra fingers, feet gliding over the pavement, mush instead of letters, objects that change, and a face that’s different by the end. Know these five glitches and you can catch them before posting and redo the shot.
Feet that skate
When a person walks, the foot carrying their weight stays still on the ground while the body moves forward. In AI video that foot often creeps along the ground as if the person were on skates. Animators call it “foot sliding”. Your eye notices it instantly, even if your brain can’t say what’s wrong. The cure is describing the step and using a reference video; we’ll cover it in level 3.
Letters and logos
Models are bad at holding text: the print on Mira’s hoodie is slightly different in every frame. In 2024 Coca-Cola released an AI-made holiday ad and got a wave of criticism. Some called it “soulless”, others said animators had to fix the logo by hand. The rule is simple: add captions and the logo in the edit from the real file, and in generation keep the fabric plain or the background blurred.
The truth about a viral short
In spring 2024 the studio shy kids released the short “Air Head”, about a guy with a balloon for a head, made with Sora. Everyone was amazed. Then the team shared details: the balloon kept changing colour, a face appeared on it, and sometimes the model drew a normal head, which they erased by hand. According to the post-production lead, roughly 1 second out of every 300 generated made the film, and the work took about two weeks. The takeaway: behind a beautiful AI video there’s almost always selection and manual polish.
Shape-shifting objects
Mira holds a cup of coffee. She walks behind a pole and comes out holding a bottle. It happens because the model doesn’t keep a list of objects in the scene: once the cup is out of view, it has to be redrawn, and it can turn into anything. What helps: keep the important object in frame the whole time, don’t hide it behind people, and keep shots short where the object gets covered.
Why exactly these break
A model learns from examples. In ordinary videos hands are small, fast and blurred, feet are often cut off by the frame edge, and the text is different in every clip. There are few clean examples, so the model has nothing to learn a rule from. The practical takeaway: the bigger and calmer the thing you care about is on screen, the fewer the errors. Want clean hands? Give them room in the frame and don’t make them do five things at once.
The 10-second check
Before showing a clip, Mira watches it at 0.25× speed and goes down the list. Hands: how many fingers, and did they merge with the cup? Feet: is the support foot planted? Face: is it the same in the first and last frame? Objects and text: did they morph? It takes ten seconds and saves you “AI slop” comments and a reshoot after posting.
Check yourself
- In the clip the model walks and her sneaker slowly glides over the pavement. What is it and what do you do?
Correct answer: Foot sliding: redo it with a step description or a reference. The support foot should stay put. If it glides, that’s a classic glitch. The cure is a precise step description or transferring motion from a real video. - Mira wants the brand name visible on the hoodie. What’s the best way?
Correct answer: Generate a plain hoodie and add the print in the edit. Models struggle to keep letters steady across frames. A print from the real file added in the edit is always crisp and identical. - A friend shows a viral AI film and says: “They made it with one button.” What’s more likely?
Correct answer: Hundreds of takes, selection and manual polish are behind it. Take “Air Head”: according to the team, about 1 second out of 300 generated made the film, and glitches like a face on the balloon were erased by hand. - Mira wants a shot of the model taking a phone out of her pocket and typing. What’s most likely to break?
Correct answer: The fingers and the text on screen. Small fast hand moves and letters are the two weakest spots of models. A close-up, slow movement and a screen replaced in the edit help.