Avoiding the AI look: making synthetic video that doesn't announce itself

There is a specific quality that makes a viewer think *AI* before they think anything else, and once that thought lands the clip is finished. Not because synthetic video is inherently bad, but because in the current moment "AI-generated" reads as a synonym for "cheap" and "low-effort." The viewer does not pause to admire the technology. They register that no one bothered to film it, downgrade their estimate of the brand accordingly, and scroll. The whole game in production right now is preventing that thought from forming. Not because hiding the method is dishonest (disclosure is a separate obligation), but because the *involuntary* tells are doing damage the disclosure label never asked them to.

The useful thing to understand is that the AI look is almost never one catastrophic glitch. It is an accumulation of small ones, each individually deniable, that together cross a threshold. The viewer cannot usually name what is wrong; they just feel that something is off, and "off" is all it takes. So the craft is not chasing a single perfect frame. It is knowing the catalogue of tells well enough to avoid the situations that produce them, and finishing the footage so the ones that slip through disappear.

The tells, and why they betray you

The current generation of models fails in recognisable places, and knowing the catalogue is half the defence:

• Hands and fine articulation. Fingers that merge, gain a knuckle, or pass through objects. Hands are the single most reliable giveaway because the model has no underlying anatomy, only statistics about how hands usually look.
• The eyes and micro-expression. A face can be flawless and still read as dead because the small involuntary movements (the saccades, the asymmetry, the way a real expression builds and decays) are absent or uncanny.
• Background instability. A scene that subtly simmers: textures crawling, straight lines wobbling, a poster on a wall that re-renders slightly differently each second. The foreground convinces and the background quietly confesses.
• Continuity drift. An object, whether a logo, a cup, or a piece of clothing, that changes shape, colour, or position across the cut. The model generates each moment plausibly but has no persistent memory of what it drew a second ago.
• Physics that is almost right. Liquid that does not quite pour, fabric that does not quite settle, weight that does not transfer. Motion the eye reads as wrong before the mind explains why.

None of these is fatal in isolation. The danger is that they compound: two or three together is enough to tip a viewer from *immersed* to *suspicious*, and suspicion is the end of the clip.

Choosing shots the model is good at

The most effective fix happens before generation, in the shot list. Models are not uniformly weak; they are weak in specific, predictable places, and a great deal of the AI look comes from asking for exactly the shots that expose those weaknesses. The discipline is to compose around the model's strengths.

That means favouring what it renders convincingly: wider framing where faces and hands are not the focus, slower and simpler motion, atmospheric and product-led shots over tight close-ups of human articulation. It means avoiding the known traps: a lingering close-up of hands manipulating a small object, a long unbroken take that gives instability time to surface, fast complex motion the physics model cannot hold together. You are not lowering your ambition. You are spending the model's competence where it has competence, and not asking it to perform in the exact places it cannot.

The fastest way to avoid the AI look is to never generate the shot that would have revealed it. Most tells are not finishing problems. They are shot-selection problems you can solve before a single frame exists.

Length is a defence

There is a simple structural reason short clips read as more real: instability needs time to surface. A background takes a beat to start crawling, a continuity error needs two moments to contradict, a not-quite-right motion needs to play long enough to register as wrong. Cut before the seams show and many of them never do.

This is why the strongest AI video tends to be quick and edited, built from a sequence of short, well-chosen shots rather than one long generated take. The cut is not just pacing; it is concealment. Every time you cut, you reset the clock on the artifacts that accumulate within a continuous shot, and you deny the viewer the sustained look that lets a subconscious doubt harden into a conscious one.

Finishing is where the seams close

The footage that comes out of a generator is a starting point, not a deliverable, and treating it as finished is one of the surest ways to look generated. Real production has always involved a finishing pass, and synthetic footage needs it more, not less. A consistent colour grade pulls disparate generated shots into one world and away from the flat, default look models tend toward. Compositing genuine elements, a real product, an actual logo, true-to-spec packaging, over generated scenes anchors the clip in something verifiably real. Grain, motion blur, and the small imperfections of a physical camera cover the unnatural cleanliness that itself reads as synthetic. And a human quality pass, watching specifically for the catalogue of tells, catches the morphing hand or the drifting object while it can still be cut.

The throughline is that avoiding the AI look is craft, not luck. It is shot selection that plays to the model's strengths, brevity that denies artifacts the time to surface, and finishing that closes the seams generation leaves open. The teams whose AI video does not announce itself are not the ones with secret access to a better model. They are the ones who treat the generator as one tool in a production process, and who do the unglamorous downstream work that has always separated footage that looks made from footage that looks cheap.

• MIT Media Lab, "Detecting synthetic media: the visual artifacts people notice," 2024.
• Stanford HAI, "Perceptual cues and the uncanny in generative video," 2025.
• Adobe, "Finishing and compositing workflows for generative footage," 2025.
• WARC, "Audience trust and the perceived effort of AI-made advertising," 2024.