When a generated clip reads as fake, it is rarely the model that gave it away. It is the choices around the model: light with no direction, a camera that never moves, and a shot held long enough for small errors to accumulate. Fix those three and the same model produces footage that passes a first look. Here is what each problem looks like and the prompt changes that correct it.
Why AI video reads as fake
Real footage carries information a viewer reads without noticing. Light comes from somewhere, so faces have a bright side and a shadow side. A camera is held by a person or mounted on a rig, so it drifts, breathes, or moves with intent. And an editor cut away before any single shot outstayed its welcome. AI video defaults to the opposite of all three: even, sourceless light; a perfectly still frame; and a clip that runs the full generation length whether the shot earns it or not. The uncanny feeling is the sum of those defaults, not a single flaw you can point at.
The good news is that all three are things you ask for, not things the model decides for you. A prompt that specifies a light source, a camera move, and an intended length is doing the work a gaffer, an operator, and an editor would do on set.
Light the shot on purpose
Flat, even light is the fastest tell. It says no lamp was placed and no window was chosen, which is a thing that never happens in real footage. Name the source and its direction and the frame gains a dimension.
Instead of "a woman in a kitchen," describe where the light comes from and what quality it has: "a woman in a kitchen, warm afternoon light raking in from a window on the left, soft shadows falling to the right." Now the model has a reason to put a bright side and a dark side on her face, and that contrast is most of what makes a face read as physical rather than rendered.
A few directions worth naming explicitly:
- Direction. Side light and backlight look filmed; flat front light looks generated. "Backlit by a low sun, rim of light along the shoulders" gives you separation from the background.
- Quality. "Soft, diffused" light for a calm interior, "hard, direct" light for a harsh exterior. The word tells the model whether shadows have sharp edges.
- Time and colour. "Golden hour," "overcast," "cold fluorescent office light." Each carries a colour temperature a viewer recognises without being told.
Give the camera a reason to move
A locked-off frame is the second tell. Real cameras are operated, and even a shot meant to be static has a small amount of life in it. Morphic lets you direct the camera move in plain language, and the vocabulary is the same one a camera operator uses. Ask for the move that suits the moment rather than defaulting to none.
1.
Pick the move that matches the intent
A slow push in (dolly forward) builds toward a subject and signals importance. A pan follows action left or right. A tilt reveals height, up a building or down to the floor. A pedestal raises or lowers the whole camera to change eye line. Say what the shot is trying to do and choose the move that says it: "slow push in on the subject's face as she realises."
2.
Set the speed and keep it small
The most common mistake is asking for too much motion. "Slow," "gentle," and "subtle" almost always beat "fast" and "sweeping." A slow push in over a few seconds reads as intent; a rushed zoom reads as a mistake. Pair the move with a pace word so the model knows the difference.
3.
Add handheld when you want it to feel captured
A touch of handheld motion is the single most humanising cue you can add. "Handheld, slight natural sway" tells the model the frame was held by a person, not bolted to a tripod. Use it for anything meant to feel documentary or intimate, and leave it off for anything meant to feel polished and locked.
A worked example: take "a car on a coastal road" and direct it as "a car on a coastal road, camera pedestals up slowly to reveal the cliff edge below, gentle handheld sway." The subject is unchanged; the move is what turns a postcard into a shot.
Cut before the illusion breaks
The third tell is length. Small errors in AI video compound over time: a hand that starts correct drifts, a background detail that held for two seconds warps by the fifth. A generated shot is usually at its most convincing in its first couple of seconds, so the edit that uses the strongest fraction of a clip beats the one that uses all of it.
Plan in short beats. Generate a clip, then take the seconds that hold and cut on motion, letting the next shot carry the story forward before the current one has time to fall apart. This is the same discipline that makes real coverage work: no shot has to do everything, because the cut is doing half the job. Assembling those beats into a single cut, with a matched audio bed, happens on Compose once you have the pieces you trust.
Put the three together
None of these fixes is exotic. Light the shot, move the camera with intent, and cut before the seams show. Applied together they move a clip from obviously synthetic to plausibly filmed, using the same model you already had. Start from a directed prompt rather than a bare description and you are already most of the way there. When you want to build a repeatable version of a look you like, the text-to-video generator is where these prompt patterns live, and the Morphic Workflows library is where you save the ones that keep working.
FAQs
The model is usually not the problem. The three most common causes are flat, sourceless lighting, a completely static camera, and shots held long enough for small errors to accumulate. All three are things you specify in the prompt rather than settings the model chooses, so naming a light source, a camera move, and a shorter beat fixes most of the "fake" feeling without changing models.
Describe the move in plain language the way a camera operator would: a slow push in, a pan left or right, a tilt up or down, a pedestal that raises or lowers the whole camera. Pair the move with a pace word like "slow" or "gentle," since subtle motion almost always reads as more real than fast, sweeping motion. You can add "handheld, slight sway" when you want the shot to feel captured rather than locked.
Name three things: direction, quality, and colour. Direction means where the light comes from, and side or back light reads as filmed while flat front light reads as generated. Quality means soft and diffused versus hard and direct, which controls how sharp the shadows are. Colour means the time and source, like golden hour, overcast, or cold office fluorescent, each of which carries a temperature a viewer recognises.
Shorter than you think. A generated shot is usually most convincing in its first couple of seconds, before small errors in hands, faces, and backgrounds have time to compound. Plan in short beats, keep the seconds that hold, and cut on motion so the next shot carries the story forward before the current one starts to break down.
Not if you keep it subtle. Gentle, slow moves and a small amount of handheld sway add realism without stressing the model. The trouble comes from asking for fast, extreme motion, which gives the model more to get wrong per frame. Match the move to what the shot is trying to do and keep the speed low.
Yes, and you should. A single prompt can name the light source and direction, the camera move and its pace, and the mood of the shot all at once. For example, "warm window light from the left, slow handheld push in on her face" carries lighting, camera, and intent together. Keep each instruction short and specific so none of them gets lost.