How Deepfake Detection Works: It's About Movement | Podcast

A deepfake can fool your eyes in a single frame. But it can't fake the way your jaw rotates across hundreds of frames. That's the difference modern detection tools actually exploit.

If you've ever worried about someone slapping your

If you've ever worried about someone slapping your face onto a fake video, this matters to you directly. YouTube recently rolled out a likeness detection tool that scans uploaded videos for A.I.-generated impersonations. Creators enroll by submitting a photo I.D. and a selfie video. The platform then uses that reference footage as a baseline, comparing it against every flagged upload. When a suspected fake surfaces, the creator gets an alert and can request the video come down. So how does the system actually tell the difference between a real face and a synthetic one?

Most people assume detection means spotting visual glitches, weird hands, broken pixels, uncanny expressions. The real method is mathematical. The system converts facial landmarks into high-dimensional vectors. Then it calculates something called cosine similarity, basically the numerical distance between a reference face and a test face. Small distance means probable match. Large distance means probable fake. It's the same principle as fingerprint analysis at a crime scene. An investigator doesn't just eyeball a smudged print. They measure ridge endpoints, whorl angles, pattern consistency. Likeness detection does the same thing with your face, frame by frame.

Why can't deepfakes just replicate that geometry? Because real faces produce behavioral biometrics that are incredibly hard to copy. Researchers have identified around sixteen distinct facial action units, things like head pitch, head roll, the horizontal distance between mouth corners, the vertical gap between your lips. From a ten-second clip, the system extracts a twenty-dimensional feature vector for every single frame. That data gets fed into machine learning classifiers trained to spot inconsistencies. Your blink pattern, the way your cheeks compress when you smile, those form a unique mathematical signature. Deepfakes were trained on datasets, not on one specific person's movement repertoire. So the geometry drifts. The micro-movements stutter. And the classifier catches it.