Why does AI video look fake?

The short version: face, physics, light, camera, sound and fine detail are the six places generated footage gives itself away, and every one of them is decided before generation. Approve the references once, write one action and one motivated camera move per shot, keep small text out of frame, and treat the sound as part of the master rather than as decoration.

This is character drift, and it is the tell people notice first. The model has no memory of who your character is: between one shot and the next it makes a fresh decision about a face it has no reason to keep.

The fix is a character bible — faces, builds, wardrobe and colour locked into one approved reference before any animation starts, and pointed at again on every shot rather than hoped for.

A model knows what an object looks like and not what it weighs. Invite it to invent the motion and it invents weightlessness: the hand that never quite grips, the cup that does not settle, the walk with no floor under it.

So the motion is written down instead of hoped for — one clear action per shot, and a short beat rather than a long one. A single action in two seconds holds; three actions in six do not.

Because nothing has decided where the light is coming from. Left to itself, every frame is lit freshly and plausibly, which is exactly what flat means: no direction, no continuity, no reason.

One approved reference sheet governs the light, the palette and the amount of detail across the whole film. Approving it once is what keeps a change of mind later a decision rather than a surprise.

Real cameras have a body, a position and a reason to move. Generated footage tends to move because movement seemed like a good idea — drifting in, drifting out, changing its mind halfway.

One motivated move per shot, decided in the shot list. Camera language is not decoration laid over the story: it is one of the ways the story is told.

Sound is the last stage of the work, and the one most often treated as decoration. A bespoke picture with a stock loop under it reads as fake in a way the picture on its own never did — the ear is quicker than the eye at catching something borrowed.

The master is where the picture, the voice and the music are fitted together. It is the stage an audience is least able to name and most able to feel.

Fine detail is where a model has the least grip: writing on a label, the pattern on a shirt, the shape of a logo. It holds for a moment and then reinvents itself.

The answer is composition rather than a longer prompt — keep the detail out of shot, hold it still, or show it once at a size the film can actually defend.

It makes each attempt prettier, and it still will not know which face you meant. A better model re-renders a wrong decision more beautifully; it does not make the decision yours.

What makes the work yours is the reference and the plan behind it, which is why all six tells have the same fix: decide it in words, approve it once, and point at that approval on every shot.

Five questions, asked before spending anything on a re-render: is the reference attached and named, have I said what cannot change, is there one action, is there one motivated camera move, and does the beat land?

When a shot fails, the answer is almost never a better prompt for the same shot. It is a shorter shot, a locked reference, or a different way of covering the beat.

The six tells have one cause and one fix, and it is not a wording trick: decide it in words, approve it once, and point at that approval on every shot. That is also why we are sceptical of waiting for the next model — a better model re-renders a wrong decision more beautifully.

If you fix one tell first, make it the sound. It is the stage most often treated as decoration and the fastest way to undo good pictures, because the ear is quicker than the eye at noticing something borrowed.

Take the prompt packs