8 min read · Updated August 2026
How to tell if a video is a deepfake
Manipulated video comes in three broad flavours, and each fails in its own way: a face swapped onto a real performance, a real face driven to say new words, and a clip generated whole from a text prompt. Knowing which one you are looking at tells you where to look.
Face swaps: watch the boundary
A swap has to blend a synthetic face into a real head. The seam runs along the jaw, hairline and ears, so pause on frames where the subject turns quickly or a hand crosses the face. Look for a soft rectangular region that stays put while the head moves, a jaw edge that flickers, or hair that briefly cuts across the face and disappears.
Skin tone at the neck is another giveaway: the face may be a slightly different colour or sharpness from the throat and ears, especially after the lighting changes mid-shot.
Lip-sync fakes: listen before you look
When only the mouth is regenerated, everything else in the frame is genuinely real — so pixel hunting outside the mouth wastes time. Watch plosives: p, b and m require the lips to close completely. Synthetic mouths often leave a visible gap.
Check that the tongue and teeth appear and disappear plausibly, that the chin and cheeks move with the jaw rather than staying rigid, and that breathing pauses in the audio line up with the body.
Fully generated clips: count the seconds and the physics
Generated video is still short. A continuous unbroken shot longer than about fifteen seconds with no cut is unusual, and objects tend to drift: a background sign changes wording, a cup changes shape, a passer-by's clothing shifts colour between seconds.
Physics is the reliable failure. Water, smoke, cloth, crowds and vehicle wheels all require consistent momentum. Look for feet that slide instead of pushing off, shadows that lag, and reflections that do not track the camera.
Audio deserves its own pass
Play the clip with your eyes closed. Cloned voices usually get timbre right and prosody wrong: even stress across a sentence, breaths in odd places or missing entirely, and no change in room sound when the speaker turns or moves away from the microphone.
Background noise is a strong tell. Real recordings have continuous ambience that ducks and swells; synthetic speech often sits on a suspiciously clean or looping bed, and the ambience may cut abruptly at the joins between generated segments.
Look for what re-uploading destroys
Every re-upload re-compresses the video, which smooths exactly the fine artefacts you would want to inspect. If you can, find the highest-resolution copy available and step through it frame by frame rather than judging a screenshot of a screen recording of a repost.
Container and stream data help here: a file whose encoder, frame rate or bitrate does not match the device it supposedly came from has been through a pipeline someone has not mentioned.
Provenance beats forensics
Before any of the above, ask where the clip came from. Is there a second angle? Did anyone else present record it? Does the account posting it have history? Was the event covered anywhere else? A verifiable original source settles a question that pixel analysis can only ever make probable.
If the claim is that a public figure said something notable, a genuine recording almost always exists in more than one place. A single copy circulating without an origin is itself a signal.
The short version
- Identify the manipulation type first; each fails in a different place.
- Plosives, jaw seams and neck colour catch most face and mouth manipulations.
- Generated clips break on physics — feet, water, cloth, reflections — before they break on faces.
- Compression hides artefacts, so always analyse the best copy you can find.