How to Tell a Deepfake From Genuine Video
There are many good indicators and few reliable ones. Each marker below works against a weak forgery and fails against a strong one — bear that in mind.
The edge of the face
The weakest point is the transition to hair, ears and neck. Look closely at the hairline: in a forgery it is often too even, as though drawn. Individual escaping strands either vanish or flicker.
Behaviour under occlusion
Watch the moment a hand, a microphone or hair crosses the face. Models handle occlusion poorly: often, for a fraction of a second, the hand ends up «under» the face, or the face loses its shape.
Fine motor detail
- Blinking: too regular, or conversely too rare
- Teeth: often a single white mass with no separations
- Tongue and the inside of the mouth: a classic site of blurring
- Reflections in the eyes: in a real person they match in both eyes
Light and shadow
Check whether the direction of shadow on the face agrees with shadows elsewhere in the frame. A forgery often preserves the reference lighting rather than the scene's. It is most visible in side light.
Why none of this is a guarantee
Every weakness listed above is a list of problems model developers are working on. Each year there are fewer. A good forgery viewed on a phone is already indistinguishable to anyone.
What works better
Not pixel analysis but provenance: where the recording came from, whether an original exists, who published it first, whether other angles of the same event exist. Context exposes forgeries more often than pixels do.
Go to the site →Read next
Face Reference: Preparing Material for a Swap Why Deepfakes Jitter on Turns, and How to Cure It Model Consent: Why There Is No Work Without It Neural Networks for Adult Content: What They Can Do TodayThis material is intended for adults over 18 and is informational in nature.