How to spot AI videos

Stay on one face for a full sentence. This is where the most work goes, so it's also where the most seams are.

• Lip sync. Mouth shapes land slightly early or late, and it shows worst on p, b, m, f, v, and w sounds.
• Teeth. They blur, merge, change shape mid sentence, or look too even and too many.
• Edges. The hairline, jaw, ears, glasses, and earrings flicker or warp when the head turns.
• Eyes. Blinking on a rhythm, a gaze that never quite focuses, reflections that don't match the room's light.
• Skin. Waxy and poreless, and it stays that way while the head moves through different light.
• Pasted on. The face is sharper, smoother, or a different color than the neck and shoulders behind it.

The photographer's note. I've photographed hundreds of faces and none of them were smooth. Real skin has pores, flyaway hairs, and a shine that shifts as the light changes. Perfection that holds still while everything else moves is generated.

Fingers got much better in still photos. In motion, the moment a hand touches something is still the hardest thing to fake.

• Fingers. They fuse, change length, or quietly change count between shots.
• Rings and nails. A ring moves to a different finger, changes color, or disappears for a few seconds.
• The grip. The hand passes through the cup, holds nothing, or wraps around an object at an impossible angle.
• The object itself. A phone becomes a slightly different phone, food changes size, a glass changes shape.

Why it works. Contact between two moving things is the hardest physics in the frame. A still image can hide it. 30 frames a second can't.

Look away from the person. The model spends its effort on the face you're supposed to watch, so the edges of the frame get the least attention.

• Text. Signs, labels, captions, and logos come out garbled, or change spelling and spacing between frames.
• Shadows. They point in different directions, or move on their own while the person stays put.
• Reflections. Mirrors, windows, sunglasses, and water show a scene that doesn't match the room.
• Physics. Hair, fabric, smoke, and water move against gravity, or with a wind nothing else feels.
• Background people. They loop the same few steps, or carry on normally through something no one could ignore.
• The camera. It drifts through a crowded space with no bumps, no collisions, and no motion blur.

Why it works. Pick 2 things that should agree, like a shadow and the light source, and check whether they do. Real scenes always agree with themselves. Generated ones almost do.

Close your eyes and listen for 10 seconds. Your ear catches things your eye forgives, and most people never think to check the audio at all.

• No breath. Real people inhale between sentences, stumble, restart, and change volume. Generated voices skip all of it.
• Metronome cadence. Even pacing, pauses in odd places, the same rise and fall on every line.
• Words. Names, places, numbers, and acronyms pronounced slightly wrong.
• The room. Studio clean dialogue on a busy street, or an echo that doesn't fit the space you're seeing.
• Mid sentence shifts. Accent, pitch, or microphone quality changing inside one breath.
• Mismatch. The emotion in the voice doesn't match the face or the situation.

Why it works. Voices get built from clean studio samples, so the messy human parts never make it in. If nothing about the delivery surprises you in 10 seconds, that's the flag.

This is the check that still works when the video is flawless, and flawless is where this is heading. It's the one to keep when you forget the other 4.

• Who posted it first. Trust the original account, a newsroom, an official agency, or the person actually involved.
• The account itself. New, anonymous, impersonating someone, or pumping out dramatic clips all day is a flag before you watch a single frame.
• What's missing. A real uploader gives you the full clip, the date, and the place. A dramatic 8 second crop with no context is doing that on purpose.
• Other angles. A genuine event produces other footage, other witnesses, and local reporting. One video and silence everywhere else is the tell.
• The official channel. If it involves a known person or company, check whether they've said anything about it themselves.
• AI labels. Platform labels and watermarks are a useful clue. They also get missed, removed, and applied to real footage by mistake, so let them inform you and keep checking.

Why it works. Visual tells have a shelf life and this one doesn't. A video can be perfectly real and still be a lie, because it's old footage relabeled as today, or a true clip carrying a false caption.