Why AI-Generated Design Looks Generic, and It Isn't the Model’s Fault

Type make me a clean, modern landing page into any of these tools and you already know what arrives. Centered hero. Gradient somewhere. Three feature cards with thin line icons. A testimonial from a person who doesn't exist. A pricing table with the middle tier highlighted.

The usual conclusion is that the model can't do design. I don't think that's what is happening, and the actual explanation is more useful to us.

You asked for the average and it gave you the average

A model returns what is most probable given what you said. When what you said contains almost no information about this particular page, the most probable answer is the center of everything it has ever seen called a clean modern landing page.

Which means generic isn't a failure state here. It's the correct answer to an underspecified question, and it would be a strange system that did anything else.

Sit with the implication for a second, because it isn't flattering. If a four-word prompt produces roughly what you would have produced, the four words were doing all the work in your process too.

What moves you off the mean

Three things, and none of them are prompt tricks.

Reference

Not adjectives. Adjectives are the problem. Clean, modern, minimal, and premium are the four most-used words in the industry and they all point at the same crowded spot on the distribution.

A specific reference does more in six words than a paragraph of adjectives. A named publication’s grid. A particular era of Swiss print. That one product you and everyone on your team already argue about. You aren't asking for imitation, you're handing over coordinates.

Constraint

Every real constraint you can state removes a mass of probable answers. One accent color. No hero image, because the client has no photography budget and never will. It has to survive being read on a phone in a warehouse. This has to work for somebody who is doing it eleven times a day and knows it better than you do.

Constraints are what made design interesting before any of this, and they turn out to be the most direct lever we have now.

The negative brief

Saying what it must not be is startlingly effective and almost nobody does it. Not a SaaS page. No card grid. Nothing that looks like it was made after 2015. You're cutting away the densest part of the distribution by name.

So where did the craft go

It didn't leave. It moved to a part of the process that used to be compressed into a few hours and is now most of the job.

The execution gap has genuinely narrowed. Making a competent thing is faster than it was and it isn't going back. What hasn't narrowed at all is knowing which thing is worth making, being able to say why this one and not that one, and having the reference library to reach for something that isn't the average.

When producing an option costs almost nothing, the scarce thing is the judgement to reject nineteen of them.

Which is the same skill it always was. It used to hide behind execution time, and now there's nowhere for it to hide.

The uncomfortable read

There's a version of this argument that flatters designers, where AI produces slop and taste saves the day. I don't think that version is quite honest.

The truer read is that a large amount of design work was already the average, and we didn't have to confront it because producing the average took two weeks and looked like effort. Now it takes nine seconds and the comparison is unavoidable.

If your work is distinguishable from the mean, this changes very little about your value and quite a lot about your speed. If it isn't, that was true before March and you had more cover.

So the thing worth building is the thing that was always worth building. A point of view, a reference library that's genuinely yours, and the willingness to say no to something that's perfectly fine.