People say they can spot AI video instantly, and they are usually right. What they are almost never right about is why.
Ask someone what gave it away and they will say the faces, or the hands. Occasionally that is true. Far more often the face was fine and the giveaway was that every shot ran exactly five seconds, or that the camera drifted forward like it was on a hovering platform, or that a busy street had no sound in it.
This is the useful part: most of the tells are not limitations of the model. They are decisions that were never made, by anyone, because the output was taken as finished when it was actually raw material. Which means they are fixable, and fixing them is what a production process is for.
Here are the ten we watch for, roughly in order of how often they are the actual culprit.
1. EVERY SHOT IS THE SAME LENGTH
The single biggest tell, and the least discussed. Most video models produce clips of around five seconds. Work that goes straight from generation to export uses each clip whole, so you get a film built from identical blocks: five, five, five, five.
No filmed sequence has ever been cut that way. Real editing gives you a two second establishing beat, a fourteen second hold, three cuts under half a second, then a long breath. Rhythm is one of the first things an audience reads, before they consciously look at anything in the frame, and evenly spaced beats read as machine output even when every individual shot is convincing.
The fix. Cut for the story, not for the clip length. Generate longer than you need and take the two best seconds. Hold a shot far past comfort when the moment earns it. Let one cut land hard on a beat and the next one sit late. If your timeline looks like a row of identical bricks, the film will feel like one.
2. THE CAMERA MOVES LIKE NOTHING PHYSICAL
Default generated camera movement is a slow, smooth, weightless push. It never accelerates, never settles, never overshoots and corrects. It is the movement of a camera with no mass, mounted on nothing, operated by no one.
Every real camera move carries evidence of how it was made. A dolly has inertia at the start and a settle at the end. A handheld frame breathes and micro corrects. A crane arcs on a fixed radius. A gimbal has that specific floaty glide with tiny step artefacts from the operator walking. Audiences have absorbed all of this from a lifetime of watching, and they notice its absence without being able to name it.
The fix. Decide what the camera is on before generating, and prompt for that specific rig with its specific behaviour. Ask for the settle. Then add the last five percent in post: a subtle handheld pass, a small speed ramp at the start of the move, weight that the model did not supply. Also, use static shots. Real films are full of frames that do not move at all, and a locked off shot is instantly more credible than a slow drift.
"The audience is not detecting a model. They are detecting the absence of anyone making decisions."
3. LIGHT WITH NO SOURCE
Generated frames often come lit from everywhere at once. Soft, even, pleasant, and impossible. Look for the shadow: if you cannot point at where the key is coming from, and the shadows on the subject do not agree with the shadows in the background, the brain files the image as artificial without ever articulating why.
Related failure: light that changes across a sequence. Shot one has hard sun from the left, shot two is overcast, shot three is back to sun but now from the right. Individually all fine, together incoherent.
The fix. Light the film on paper before generating anything. One key direction, one time of day, one quality of light, written down and applied to every prompt in the sequence. Ask for hard light more often than feels safe, because hard light has a direction and directional light reads as real. Then match the shots to each other in the grade, which is where the last inconsistencies get resolved.
4. SKIN AND SURFACES WITH NO TEXTURE
Models tend toward clean. Skin without pores, without asymmetry, without the small blemishes and shine that make a face look like a face. Walls with no scuffs. Fabric with no wear. Metal with no fingerprints.
This is the uncanny quality people describe as plastic, and it is not really about the face at all. It is about the world being too new. Everything real has a history and shows it.
The fix. Ask for imperfection explicitly at still stage: visible skin texture, flyaway hair, worn surfaces, dust, fingerprints on the glass. Then in the grade, reintroduce what a sensor would have produced. Real grain, not a digital noise overlay. Slight chromatic aberration at the edges. A touch of highlight rolloff instead of clipped whites. Halation where a bright source blooms. These are the fingerprints of physical capture, and putting them back is one of the highest impact things you can do in ten minutes of colour work.
5. THE FACE CHANGES BETWEEN SHOTS
The tell everyone knows. The same character across a sequence, but the jaw is slightly different, the hairline moved, the age drifted by five years, the shirt changed shade.
What makes this so damaging is that the audience does not consciously catch it. They just stop believing in the person, and then in everything around them.
The fix. This one is won before generation, not after. Lock the character as an approved still, with reference images, and carry that reference through every shot in the film. Approve all the frames of a sequence side by side rather than one at a time, so drift is visible while it is still cheap to correct. This is exactly why still frame approval sits at step four of our pipeline, before anything moves. Also, direct around the problem: fewer, longer shots of a character beat many short ones, and a figure seen from behind or in profile carries far less consistency risk than a full frontal close up.
6. HANDS, FEET AND CONTACT POINTS
Hands have improved enormously and are no longer the reliable giveaway they were. What still breaks is contact: the point where one thing touches another. Fingers gripping a glass that has no weight. Feet meeting the ground without settling into it. A bag strap resting on a shoulder without deforming either. Objects passed between hands that briefly merge.
The physics of contact is the hardest thing for these systems, and it is also where a viewer's attention naturally goes.
The fix. Direct the shot to avoid the hardest moment. Show the hand already holding the glass rather than the act of picking it up. Cut away at the instant of contact and back after it. Frame out the feet on a walking shot. This is not evasion, it is the same shot selection logic that has always governed how directors handle expensive or difficult action. When contact is essential to the story, generate many takes and accept that most will fail.
7. BACKGROUNDS THAT WILL NOT HOLD STILL
Watch the edges of the frame rather than the subject. People in the far background walking through each other or changing clothes mid stride. A parked car that becomes a different car. Building windows shifting their grid. Signage that resolves into letters that are not letters.
Text is the worst of it. Arabic and English signage in a generated street scene is very frequently nonsense, and in a Dubai or Riyadh setting a local audience reads that instantly.
The fix. Keep backgrounds simple and shallow. A wide open depth of field with a hundred moving details is asking for trouble; a long lens with a soft background is both more cinematic and more stable. Any text that matters gets composited in afterwards as a graphic, correctly typeset, rather than trusted to the model. And check the corners of every frame during review, because that is where the failures live and where nobody looks.
8. THE DEFAULT PALETTE
There is a look that a lot of this work shares: rich teal shadows, warm orange highlights, high contrast, high saturation, everything slightly glossy. It is attractive on a single frame and it is the visual equivalent of a stock font.
The reason it recurs is that it is what these systems produce when nobody specifies otherwise, and it is what the training data rewards. When your film shares a palette with ten thousand other films made the same month, the audience recognises the family even if they cannot name the relative.
The fix. Commit to a palette in art direction and enforce it in the grade. Pull saturation down further than feels comfortable. Choose an unfashionable colour relationship. Let the shadows go properly black or properly milky instead of teal. A distinctive grade is the cheapest way to make work look authored rather than produced.
9. NO SOUND, OR SOUND THAT IS ONLY MUSIC
A shot of a busy souq with nothing but a music bed under it will never feel real, no matter how good the picture is. Half of the credibility of a filmed sequence comes from the ears, and generated work is very often delivered with a track dropped on top and nothing else.
Real footage carries room tone, the specific acoustic signature of a place. It carries footsteps, fabric, traffic, the ambient wash of a room. Remove all of that and the eye starts hunting for other reasons to be suspicious, and it finds them.
The fix. Build a full sound design. Room tone under everything. Foley for every significant movement in frame. Ambience that matches the space, and that changes when the location changes. Where there is dialogue, cast a voice that belongs to the region rather than a generic accent, and mix it into the space rather than on top of it. Sound is the single cheapest credibility upgrade available and it is skipped more than any other step.
10. SHOTS THAT NO PRODUCTION WOULD EVER BUILD
The subtlest tell. A shot that is technically flawless and makes no production sense: an impossible crane through a solid wall, a scale that no location provides, a set that no budget would build, five completely different environments in a fifteen second spot for a small brand.
Viewers carry an unconscious sense of what things cost. When the images imply a budget the brand obviously does not have, the conclusion is not that the brand got lucky. It is that none of this is real, and by extension neither is the claim being made.
The fix. Direct within a plausible production. Choose shots a crew could have got. Keep the environment count honest for the format. Use the impossible shot once, deliberately, as a moment, rather than as the default vocabulary of the whole film. Restraint here is what makes the one big shot land.
THE PATTERN UNDERNEATH ALL TEN
Read the fixes back and something becomes obvious. Almost none of them happen at the moment of generation.
Shot length is an edit decision. Camera weight is direction plus post. Light is art direction. Texture and palette are grading. Character consistency is frame approval before anything moves. Sound is its own discipline entirely. Two of the ten are prompt problems. The other eight are the parts of filmmaking that existed long before any of these tools did.
That is the whole argument for treating this as production rather than generation, and it is why quotes for the same brief vary by a factor of thirty across this market. A studio selling generation cannot fix nine of these, because it has no editor, no colourist and no sound. A studio running a full pipeline fixes them as a matter of routine, and the work stops announcing how it was made.
Which is the actual goal. Not to hide that AI was used, we are open about it and clients are increasingly relaxed about it. The goal is that the audience is thinking about the brand rather than about the tool.
