An ad came up on Instagram recently for a well-known Indian appliance brand. The product itself looked good: cleanly lit, sitting properly in its own shadow, holding the frame. That part was clearly 3D. Everything around it was clearly not.

Elements drifted through the shot in positions nothing should occupy. They moved without weight. One of them changed into something else halfway through its own travel, and the ad carried on as though nothing had happened. If you know what generative video does when it is left to fill in the gaps, it took about two seconds to read.

The interesting part is not that the AI looked bad. Some of it looked rather good. The interesting part is that the ad stopped being believable anyway.

The tell was not ugliness

The failure mode of generative video is not poor quality. It is drift. A shape changes identity between frames. An object moves as though it has no mass. A surface that was brushed metal a moment ago is now something closer to plastic, and nothing in the edit acknowledges it.

Most viewers will never name any of this. Nobody watching an appliance ad is thinking about temporal coherence. They simply feel that something is off, and they stop extending the frame the benefit of the doubt.

That doubt does not stay where it started. Once one element in a shot reads as fabricated, everything sharing the frame inherits the suspicion, including the parts that were true. As far as I could tell, the product in that ad was represented honestly. It did not matter. It was keeping bad company.

A 3D model is a reference point. A prompt is not.

This is the whole argument, and the rest follows from it.

A 3D model is built once, to measurement, and it then sits there being exactly itself. Every frame is a camera looking at that same object. It cannot drift between shots because nothing re-decides it between shots. Move the camera, change the lens, relight the scene, cut to a close-up of the control panel, and the geometry underneath is the same geometry it was in the wide.

Generative tools do not work this way. Each frame is produced afresh. Reference images, seeds and conditioning all narrow the variation, and they have narrowed it a great deal in the last two years, but they do not eliminate it, because what is being asked for is a new image resembling the previous one rather than the same object seen again.

That sounds like a technical distinction. It is really a commercial one. 3D can promise you the same object twice. Generation, as it currently stands, can promise you something close.

The product is the claim

In product advertising the product is not decoration. It is the promise being made.

The number of vents on the guard. The proportion of the base to the column. Where the badge sits and how big it is. The exact colourway, which marketing spent a month choosing and the factory spent longer matching. These are not details a viewer consciously audits. They are the thing being sold, and somebody is deciding whether to spend money on the strength of that image.

Generative tools have no concept of this particular product. They have a concept of an appliance roughly like this one. Ask for another angle and you will get something plausible, and plausible is precisely the problem: whatever it invents becomes a claim your brand did not make, cannot stand behind, and is showing to somebody at the moment they are deciding to buy. The same logic applies to what goes on the pack, where the gap between the image and the object in the customer’s hand is even shorter.

Where AI earns its place

None of this is an argument against the tools. I use them.

They are genuinely good at the parts of the process where nothing is being claimed. Ideation. Mood boards. Look development. Quick roughs that let a client react to a direction before anyone commits real hours to it. Environments and abstract backgrounds that exist to hold a mood rather than to describe a thing you can buy. Used there, they save days and cost nobody anything.

The line I work to is straightforward. Use AI where nothing is being claimed. Use 3D where the product is the claim. That is close to the distinction I drew about writing in an AI world: the tools are useful right up to the point where the output has to carry the brand’s word.

The 3D model is an asset you keep

There is also a practical argument that has nothing to do with trust.

A 3D model is something a brand can own, lock down and change at will. The same one lights the film, the listing images, the carton mock-up, the buyer’s deck and next year’s festive variant, and it does all of that at whatever size the shelf happens to be. Build it once and you can shoot it for years, in a week, without booking anything.

A generated image is a one-off. You cannot return to it and ask for the same thing from a different angle, because there is no thing to return to.

The working relationship is different too. A 3D model takes direction: rotate it fifteen degrees, drop the key light, make the cap matte, show me the version with the second colourway. Those are instructions with predictable outcomes. Generation takes re-rolls, where you regenerate and hope, and the detail you liked last time may not survive the attempt.

What the customer is owed

Underneath all of it is something simpler than any of the above.

People extend a certain trust to brands they recognise. That trust takes years and a great deal of money to build, and an ad showing a product that does not quite exist spends it to save a few days of production.

The audience able to spot this is growing, not shrinking. Every month more people learn the tells, because more people are using the tools themselves. What passes as a clever shortcut this year will read, fairly soon, as a brand that could not be bothered to show you the actual thing.

Customers deserve to be able to trust what they are looking at. Brands they already know should be the last ones to make that a question worth asking.

Launching something and want it to look like itself? Let’s talk →