How to detect deepfake videos

Most deepfakes are strongest when the face is frontal and well-lit. Profiles, sharp turns, and partial occlusions (a hand passing in front of the face) are where the fake breaks. Look for the jawline shifting subtly, the ear changing shape, or the eye line drifting off the head.

Real speech has tight phoneme-to-mouth alignment. Plosives (P, B), bilabials (M), and rounded vowels (O, U) all involve specific lip closures or shapes. Deepfake lip-sync is improving but still tends to over-smooth these moments. Listen for a 'P' while watching whether the lips actually close.

If the audio itself was cloned, the cadence often gives it away — too uniform, missing the small breath gaps real speakers leave between phrases.

Face-only deepfakes leave the rest of the scene untouched. The result: the face has slightly different compression than the background, the lighting on the cheek doesn't match the lighting on the shoulder, and the edges around the jaw flicker between frames.

Pause the video. Step frame-by-frame. Watch the face boundary, not the eyes.

The cheapest deepfake is no deepfake at all — a real video with a misleading caption. Before you score the pixels, score the source. Who posted it first? Does the original account have a history? Does the claim match other reporting?

Automated detectors look at facial-landmark stability, lip-sync alignment, and scene continuity at scales humans can't track frame-by-frame. A score is not a verdict, but combined with eyes-on review and source verification it changes the calculus on whether to share.