What AI is genuinely good at in ad creative, and where it still falls over

Generative image models have been oversold and, in a narrower way, undersold. Oversold as a replacement for creative judgement. Undersold as a solution to the specific, boring bottleneck that stops most small teams from testing enough creative: producing the fifth variation.

Having built an ad generator, here is the honest split of what works and what does not.

What they are genuinely good at

Volume without fatigue

The tenth variation costs the same as the first. That single fact changes what testing is affordable. A team that previously shipped one execution per campaign because that was the budget can ship eight and let performance data pick the winner. This is the real value, and it is not glamorous.

Backgrounds, textures and abstract composition

Diffusion models are excellent at the material qualities of an image: lighting, depth, atmosphere, surface. Ask for "soft studio light on a matte surface with a shallow depth of field" and you get something usable immediately. This is a real substitute for stock photography, which is expensive and looks like stock photography.

Breaking a blank page

Even output you reject is useful, because reacting to something is far easier than starting from nothing. Generating six directions and immediately knowing five are wrong tells you a great deal about the sixth.

Where they still fall over

Text. Specifically and reliably, text.

This is the failure that matters most for advertising, because an ad is mostly a sentence. Diffusion models generate text as a texture that resembles writing rather than as writing. You get plausible letterforms arranged into words that are subtly wrong: a doubled character, a dropped one, kerning that no typographer would allow.

Brand consistency across a run

Ask for the same brand twice and you get two different interpretations. Without an external constraint (an actual palette, an actual logo, an actual layout system), output drifts. This is why a generator that starts by reading your live site produces more usable results than one that starts from a text prompt alone.

Knowing what your product does

A model has no idea whether your claim is true, whether it is the claim worth leading with, or whether it is compliant in your market. It will happily generate a beautiful advertisement for something you cannot legally say. Judgement about the message is not automatable and should not be.

The practical division of labour

1. You decide the message: the one idea worth putting in front of a stranger.
2. The model produces the visual, at volume, with variation you would not have had budget for.
3. A rendering engine sets the typography, correctly, at native platform sizes.
4. You judge the output and kill most of it, which is fine, because generating the next batch is cheap.

Framed that way, generative tooling is not replacing a designer. It is removing the part of the job that was never design in the first place: producing the eighth size of the third variation before Friday.

That is a narrower promise than the marketing around this technology usually makes. It is also the one that survives contact with a real campaign.