Published on 2026-07-25
What can't a deepfake fake, at least for right now?
Pass a hand across your face on a video call and a real-time fake often falls apart. That trick works today. It is not built to last, and here is why.
Here is a strange party trick with a serious use. On a video call you have doubts about, ask the person to slide a hand slowly across their own face. Watch closely. A live, generated face will often glitch, smear or lose its shape right where the hand passes over it, because facial occlusion, anything blocking part of the face mid stream, is one of the places synthetic video still struggles to keep up convincingly.
The moment something gets in the way
The reason is almost mechanical. These systems build a face by tracking and reconstructing it frame by frame in real time, and a hand crossing that face interrupts the tracking exactly when the software needs it most. Sticking out a tongue causes a similar problem, along with plain old lip to speech synchronization, which real time systems still have to compute on the fly rather than polish frame by frame the way a pre rendered video can be polished.
This matters most on a live call, where nobody gets a second take. A prerecorded fake can be rendered, checked, rerendered and only released once the seams are smoothed over. A generator running live during an actual conversation has no such luxury: it has to guess what your hand and your face will do next while you are already doing it, and an unscripted, ordinary gesture is exactly the kind of thing that is hardest to predict in advance.
Turn your head and watch the seams
A second, well documented weak spot is the side profile. The software typically used to map a face onto footage relies on facial landmarks, reference points around the eyes, nose and jaw, and testing has found that a profile view gets assigned roughly half to two thirds as many of those landmarks as a straight on shot does. Push the angle to a full 90 degrees and the image visibly distorts. Part of the problem is simply data: profile photos are rare online compared with the flattering, front facing shots most of us post, so there is less material for any system to learn from.
- A hand or object passing in front of the face, which can smear or glitch mid motion
- A full side profile, especially near a 90 degree angle
- Sticking out a tongue, or exaggerated mouth shapes
- Fast, complex, full body movement rather than a calm head and shoulders shot
A snapshot, not a rulebook
Here is the part worth remembering longer than any single trick on that list: none of it is a law of nature. Siwei Lyu, a computer scientist at the University at Buffalo who has spent years researching deepfake detection, points out that the flicker, warping and distortion around the eyes and jawline that used to reliably expose a fake have largely vanished from today's best generators. 'Simply looking harder at pixels will no longer be adequate,' he says of where detection is heading. Every weakness researchers publish becomes a bug the next version quietly fixes, which means a checklist like this one is a photograph of this month, not a permanent map.
A trick that catches this year's fake is a description of this year's software, nothing more.
What actually holds up over time
If the visible tells keep expiring, the sturdier habit is to stop treating the video itself as the evidence. Lyu's own answer points toward infrastructure rather than eyesight: media signed cryptographically at the moment of capture, so a file can prove where and when it genuinely came from, and forensic tools that combine several signals instead of betting everything on one glitch. You do not need the cryptography yourself to borrow the underlying instinct. Where did this video first appear, who else is showing it, does the original source confirm it happened the way the clip claims? Those questions do not age the way a hand test does. They worked a decade ago, on grainy edited photos, long before anyone said the word deepfake out loud, and they will still work on whatever comes after this generation of software.
So the hand test, the profile turn, the tongue, they are worth trying, and worth teaching to anyone who asks you how to tell. Just hold them loosely. The face on your screen is being generated faster and more convincingly with each passing season, and the one thing that has not changed is where the real proof has always lived: not in the pixels, but in where the video came from and who else can confirm it.
Sources
- Deepfakes leveled up in 2025: Here's what's coming next (University at Buffalo, UBNow) (buffalo.edu)
- Face biometrics' limitations with profile images could help deepfake detection (Biometric Update) (biometricupdate.com)
- Can You Spot a Deepfake? Are You Sure? (Corporate Compliance Insights) (corporatecomplianceinsights.com)