AI video generators have improved dramatically in two years. Hands are better. Backgrounds are cleaner. Voices are harder to distinguish from real speech.
The face still breaks in predictable ways.
Not every AI face breaks in every video. But the failure patterns are consistent enough that you can build a fast mental checklist and run through it in under thirty seconds.
The Face Is Where AI Video Still Loses in 2026
These six tells show up across generators, across resolutions, and across content types. Check them in order. The first two are the fastest to spot.
The list is a starting point, not a guarantee. A video can pass all six and still be synthetic. Use these as signals, not verdicts.
1. The Ear-Jaw Boundary
The junction where the ear meets the jaw is where most face generators lose coherence. Look at the area between the earlobe and the chin. On a real face this is a clean anatomical line. On an AI face it is often slightly blurred, smeared, or geometrically inconsistent with the angle of the head.
Earrings make this easier to catch. A real earring moves with the earlobe. On AI-generated faces, earrings sometimes float slightly off the lobe or lag behind head movement by a frame or two.
2. Teeth During Open-Mouth Speech
Teeth are a consistent failure point. AI generators either produce teeth that look too uniform, every tooth identical width, identical length, identical brightness, or teeth that blur and lose definition when the mouth opens wide.
Watch for a pause in the video where the mouth is fully open. In that single frame, look at whether the teeth have individual texture and variation. Real teeth are imperfect. AI teeth tend toward either a perfect white row or a smeared approximation.
3. The Philtrum in Motion
The philtrum is the vertical groove between the nose and the upper lip. On a real face it deforms naturally when the person speaks, smiles, or raises their upper lip. On many AI faces it stays rigid while the mouth moves around it. Or it deforms in a way that is geometrically inconsistent with the surrounding facial muscles.
This is easiest to catch mid-sentence when the person is speaking quickly. Slow the video to 0.5x speed in the TikTok player and watch the nose-to-lip area during fast speech.
4. Eye Moisture and Catchlights
Real eyes have moisture. That moisture creates a small, bright catchlight reflection of the light source in the corner of the eye. AI-generated eyes either have no catchlight, have a catchlight that does not correspond to any light source in the background, or have a catchlight that stays fixed as the head moves.
On a real face the catchlight shifts slightly as the person turns. It reflects the actual environment. On a generated face it is often painted on as a static element.
5. Skin Texture Under Motion Blur
Skin has texture: pores, fine lines, slight variation in tone. AI-generated skin handles this inconsistently in motion. In a static frame it can look convincing. During motion, particularly when the face turns or the camera pans, the texture either smears or flickers. The texture is regenerated frame-to-frame at a slightly different scale or position.
Watch the forehead and cheeks during any head movement. Real skin texture is continuous. AI skin texture often shimmers very slightly at the edges of motion.
6. The Neck-to-Shoulder Transition
Face generators are trained primarily on faces, not on the full head-neck-shoulder system. The seam where a generated or swapped face meets the original neck is where generation artifacts concentrate.
Look at the collar area and the sides of the neck. On face-swap deepfakes, there is often a subtle color or texture mismatch at the boundary. On fully synthetic people, the neck and shoulders sometimes have a slightly different rendering quality than the face above them.
Visual Detection Has Real Limits
These tells apply most reliably to lower-quality or older generation models. The best current generators handle some of these artifacts better than others. Lip sync and teeth are the most persistent failures. Eye behavior and skin texture are improving fastest.
A video can clear every check on this list and still be AI-generated. This is not a reason to stop looking. It is a reason to check the Ledger community record alongside your own visual assessment. A face that passes a quick visual check may still be from an account that Ledger users have already flagged through repeated encounters.
Visual inspection and community reporting are not alternatives. They work together.
Train Your Eye to Spot AI-Generated Faces
Now that you know the tells, practice spotting them. Ledger's Train Your Eye mode shows you real and AI-generated video clips and scores your accuracy. The more you train, the faster you catch what slips past a casual scroll.
Start training at ledgerapp.app/play
Related Posts
- How to Tell If a TikTok Video Is AI-Generated: 7 Signs to Check Right Now: the full platform-specific detection guide for TikTok
- What Is a Deepfake? A Plain-English Guide for Social Media Users: what the different types of AI-generated video actually are and how they are made
- What to Do When You Find a Deepfake on TikTok or Instagram: the step-by-step action guide for after you have spotted something suspicious
