Why Your AI Video Looks Fake: 9 Tells and How to Fix Each One

Published October 2026 · 12 min read

You generated a clip. Technically it worked — the render finished, the subject is recognisable, nothing melted into a nightmare. And yet within two seconds a viewer knows it is not real. They cannot always say why. They just know.

That gap between “technically generated” and “believable” is almost never the model's fault alone. Modern video models are extremely literal: they render exactly what you describe and invent the rest. When the invention is wrong, you get the uncanny result. The fix is usually not a better model but a more complete description.

Below are the nine tells we see most often in our own work, in rough order of how quickly they break the illusion. For each one: what it looks like, why the model does it, and the specific wording change that helps.

1. Skin that looks like moulded plastic

What you see: faces with no pores, no fine lines, no variation in tone. The surface is uniformly smooth, sometimes slightly waxy, and it catches light the way a mannequin does.

Why it happens: models are trained on an enormous volume of retouched, beauty-filtered imagery. Ask for a person and the statistical average of “person” in the training data is a person who has been through a retouching pipeline. Words like beautiful, professional or commercial push it further in that direction.

The fix: describe imperfection explicitly. Texture words do more work than quality words.

close-up of a 40-year-old man, weathered skin, visible pores, faint stubble, small scar above the left eyebrow, uneven skin tone

Drop beautiful, perfect skin, flawless. They are actively harmful. If you need an attractive subject, describe the features you want instead of the verdict — “high cheekbones, clear eyes, strong jaw” rather than “beautiful woman”.

2. Eyes that do not track

What you see: in a held close-up, the pupils drift slightly, the iris changes shape, or both eyes fail to converge on the same point. Sometimes the eyelids move at a different rhythm from the rest of the face.

Why it happens: eyes are the highest-information part of a human face and the hardest thing to keep coherent across frames. Any small inconsistency that the brain would ignore on a hand becomes unbearable on an eye. Long, static close-ups give the model the maximum number of frames in which to make a mistake.

The fix: two options, and both work. Either avoid the sustained close-up, or give the eyes something specific to do.

medium shot, subject looking slightly off-camera to the left, eyes tracking a passing car

Directional gaze is much more stable than a blank stare into the lens. If you genuinely need a locked-off close-up, shorten the shot to two seconds and cut away before the drift becomes visible.

3. Motion with no weight

What you see: a person walks but appears to glide. Feet do not plant. A heavy object is lifted with the effort of picking up a cup. Cloth hangs as though the body inside it is not moving.

Why it happens: the model has learned the visual pattern of walking without necessarily modelling mass, friction or ground contact. It reproduces the shape of motion, not its physics.

The fix: name the physical consequences. Ground contact, weight shift and effort cues are all describable.

man in heavy boots walking on wet gravel, each footstep compressing the surface, slight shoulder dip with each step, weight shifting visibly

Low camera angles help because they make foot contact visible. Wide shots where the full body is in frame and the ground is in shot also help — the model has more context to keep the feet honest. A cropped waist-up shot of someone “walking” is where floating is easiest to get away with and easiest to spot.

4. Hands

What you see: too many fingers, joints bending the wrong way, fingers merging into an object, or a hand that changes shape between frames.

Why it happens: hands are small, highly articulated, frequently occluded, and rarely the subject of the shot — so they receive less of the model's attention budget. They are also the classic failure case that every model has been specifically pressured to improve, which means results vary wildly between versions.

The fix: do not fight it, stage around it. Hands in pockets, behind the back, holding an object that defines the silhouette, or out of frame entirely.

man with both hands in his jacket pockets, shoulders relaxed

If the shot genuinely requires hands doing something, keep them moving and keep the shot short. A hand in motion reads as fine; a hand held still in the foreground gets scrutinised frame by frame.

5. Hair and fabric with no inertia

What you see: a subject turns their head but the hair arrives at the same time. A coat moves as a rigid shell. A scarf hangs vertically while the person runs.

Why it happens: secondary motion is expensive to model and easy to omit. If you do not mention it, the model often renders hair and cloth as attached geometry rather than as material with its own momentum.

The fix: describe the material and the motion it should be reacting to.

woman turning sharply, long hair swinging a beat behind the head, loose linen shirt catching the wind

Specifying the fabric (“stiff denim” versus “loose linen”) also sets the expected amount of movement. Heavy fabric should move less; light fabric should move more. Getting that relationship right is a subtle but strong realism cue.

6. Lighting that belongs to another room

What you see: the subject is lit from the front but the background shadows fall to the left. Two light sources with different colour temperatures. A face that is evenly lit while standing in a dark alley.

Why it happens: lighting is described separately from the scene unless you connect them. The model renders a nicely lit subject and a plausible background, and the two are not always reconciled.

The fix: state the light source, its direction, its quality, and its consequence.

single hard window light from the left, strong shadow falling across the right side of the face, dust visible in the beam

Naming a physical source (window, streetlamp, phone screen, overcast sky) is more reliable than naming a mood (“cinematic lighting”). A mood word gives the model latitude to invent; a source word constrains it. Our lighting phrase library has a longer list of source-based descriptions.

7. Background crowds

What you see: faces in the background that are smeared, duplicated, or subtly wrong. People who walk into each other. A crowd where everyone is facing the same direction.

Why it happens: background figures get fewer computational resources and less prompt attention. When there are many of them, each individual gets almost none.

The fix: reduce the number and increase the blur.

empty street at dawn, shallow depth of field, background heavily out of focus

Three or four background figures read as a crowd. Twenty read as a rendering error. If the scene demands density, shoot it as a silhouette or in heavy rain or fog — atmosphere hides exactly the detail the model cannot produce.

8. Signage, text and logos

What you see: shop signs with letters that are almost right. A t-shirt with a word that dissolves on closer inspection. A phone screen showing gibberish.

Why it happens: rendering legible text requires character-level accuracy that video models are not built for. The result is text-shaped texture, not text.

The fix: keep text out of the frame, or accept it as an unreadable blur, or add it in post.

storefront with unmarked frosted glass, no signage

If a sign is essential to the story, generate the plate without it and composite the sign in your editor. It takes two minutes and it is the difference between a shot that survives scrutiny and one that does not.

9. Camera moves that break physics

What you see: a push-in that accelerates and decelerates with no reason. A orbit that changes radius mid-move. A handheld shot that is shaky in a way no human arm shakes.

Why it happens: complex compound moves are described in words but executed without an internal model of a camera rig. The model interpolates between the start and end states and the middle is whatever is statistically plausible.

The fix: use one simple move per shot and describe its speed.

slow steady push in, constant speed, tripod-stable

Single moves — push in, pull out, pan, tilt, static — are far more reliable than combinations. If you want the energy of a complex move, cut between two simple ones. Our camera move reference lists which ones hold up and which tend to fall apart.

The pattern behind all nine

Look back at the fixes and a single principle emerges: every failure is the model filling a gap you left open. You did not say the skin was textured, so it defaulted to the average of retouched imagery. You did not say which direction the light came from, so it guessed. You did not say the hair should lag, so it did not lag.

This is why adding more adjectives rarely helps and adding more specifics usually does. “Cinematic, beautiful, high quality, 4K” adds no information the model can act on. “Hard window light from the left, 40-year-old man, weathered skin, hands in pockets” adds four constraints that eliminate four failure modes.

A quick pre-flight checklist

Before you spend credits on a shot, run through this list. Each line maps to one of the tells above.

Where to go next

If you want to skip the manual assembly, the prompt generator walks through these same fields in order — subject, texture, lighting, camera, motion — and assembles a prompt you can paste directly. The negative prompt reference covers what to exclude when a specific failure keeps recurring, and the 12 costly mistakes guide covers the budget side of the same problem: the retries that these tells quietly cause.

One last thing worth internalising: none of this is permanent. Each of these failure modes has been steadily improving across model generations, and some of them — hands especially — are already much better than they were a year ago. What does not change is the underlying principle. The model renders what you describe. The craft is in what you choose to describe.