Where AI in video makes sense, and where it is a trap

Over the past two years the line between what is worth filming and what is not has moved. A setting that would require a flight, a product that does not exist yet, twenty language variants of one visual — all of it can now be made without a crew.

At the same time a category of mistakes has appeared that used to be impossible: a photograph of an employee who never worked at the company, and a satisfied customer who does not exist. That is not a technical problem. It is a decision.

Where it works

Generated imagery is good for what cannot be filmed, or cannot be filmed economically.

An unfinished product. Visualising packaging still in production, or a machine still on a drawing. The classic alternative is 3D visualisation — just slower and more expensive.

A setting beyond the budget. A backdrop that would mean flying abroad, or a space a crew cannot enter. When the real product and real people are in the foreground, a generated background misleads nobody.

Variants. A campaign needing one visual in twenty guises for different markets and formats. Filming twenty variants means twenty setups; generating them means an afternoon.

Completing a shot. Extending the frame to another aspect ratio, removing a distraction, stretching a scene by the second the edit is missing. This is routine post-production today and nobody calls it AI.

Where it does not

People meant to read as your employees. It is not only that it shows — hands and eyes remain the weak point. It is what happens when someone notices. A brand caught using a generated photo of “our team” saves once and pays for it for a long time.

A product the customer will buy and compare. Generated images of food, clothing or interiors look better than reality. That is the definition of misleading advertising, and it gets settled by a regulator rather than by a conversation about how interesting the technology is.

References and evidence. A photo from an event that did not happen; a quote from someone who said nothing. At that point it is no longer marketing.

Labelling is ceasing to be voluntary

European AI rules introduce an obligation to label artificially generated or manipulated content that resembles real people, objects or events. Platforms add requirements of their own, and some already label automatically from metadata.

The practical consequence for advertisers: you need to know what in your material is generated — and the cheapest way to know is to write it down as it is made. So with every job we deliver a breakdown of which shots were filmed and which were generated. Reconstructing that a year later is work nobody wants to do.

What about copyright

The generated output itself is not reliably protected as an authored work in Czechia or the EU — without a human creative contribution, no protected work arises. In practice that means contracts cover a licence to use rather than a transfer of ownership, and generated parts are combined with filmed material, which is protected.

The simple advice that follows: for content meant to be exclusive — a campaign key visual, an ident, a mascot — it is safer to have something filmed or drawn at its foundation.

The practical conclusion

The best results tend to be mixed. A filmed hero shot with the real product and real people, a generated backdrop or a completed scene. The viewer gets a credible image, the budget carries one shoot day instead of three — and when the brief changes, only the generated part changes.