Why AI Video Looks Uncanny — and the Six Fixes That Work
The uncanny look is not one problem. It is six specific failure modes with six specific causes, most of them fixable in the prompt or the input rather than by paying for a better model.
"It looks AI" is a verdict, not a diagnosis. In practice it decomposes into a handful of distinct artefacts, and viewers register them in a predictable order. Each has a different cause and most are fixable without changing model.
1. The dead face
The most common one. Features are technically correct but nothing moves — no micro-expression, no blink variation, no weight shift.
Usually a prompt problem. A prompt describing a state ("a woman standing in a kitchen, smiling") gives the model nothing to animate, so it animates nothing. Describe a sequence: "first she looks down at the cup, then raises her eyes to camera and smiles." Actions in order beat adjectives every time.
2. Drift in the second half
The face is right at second one and subtly wrong at second eight. Wardrobe shifts, hair changes length, features soften.
Drift compounds with duration and with motion amplitude. The fix is not a better prompt — it is a shorter clip. Five to eight seconds is where output is reliably stable. If you need fifteen, generate two shots and cut; a cut is invisible and drift is not.
3. Hands and contact points
Hands remain the giveaway, along with anywhere two things touch: fingers on a cup, a strap on a shoulder, feet on the floor.
Partly a model limitation, partly framing. Keep hands out of the tightest part of the frame when you can, and prefer clips where contact points are stable rather than constantly changing.
4. Wrong physics
Hair that moves like a solid, cloth that ignores momentum, a body that does not settle its weight. Viewers cannot name this one, but they feel it.
This is genuinely model-dependent. Newer motion models simulate cloth and hair dynamics rather than pasting motion onto a rigid rig, and it shows. If everything else is right and the clip still feels wrong, this is where paying more actually buys something.
5. The lighting mismatch
Common when a person was generated from reference photos: the subject is lit one way and the scene another, so they read as pasted in.
Fix it in the input. Supply reference photos with even, neutral lighting, and name the scene's lighting explicitly in the prompt so the model has one instruction rather than two conflicting ones.
6. Too smooth
Perfect skin, no grain, no lens character, unnaturally steady framing. Real phone video is noisy and slightly imperfect, and its absence reads as synthetic even when nothing is technically wrong.
Ask for the imperfection: handheld framing, natural skin texture, practical light sources.
The ranking that matters
If you only fix two things, fix clip length and prompt structure. Shorter clips and action-sequence prompts remove more of the uncanny look than any model upgrade, and both are free.
Per-second prices are close enough across the credible models — roughly $0.10 to $0.17 — that switching model is rarely the cheapest lever. Re-running a five-second clip costs under a dollar. Re-running a fifteen-second one three times because it keeps drifting costs five, and still drifts.