Scroll through the apps built with AI this year and you'll notice something: they all kind of look the same. That's not a failure of taste, it's a property of how these models work. And the way out isn't a better prompt. It's architecture.
Once you see it, you can't unsee it
Put a dozen AI-generated apps side by side.
A hero section with a centered headline. Card grids with rounded corners. A blue-to-purple gradient somewhere. Inter, or something close to it. Soft shadows, generous spacing, the same style of empty-state illustration. Different products, different companies, different prompts. One aesthetic.
None of it is ugly. Most of it is competent. The problem is elsewhere: your app is supposed to be yours, and it looks like everyone else's. On a store listing, the design was your one chance to say "this is us" before anyone reads a word.
People call this AI slop. The name is harsher than the output deserves, most of these screens are fine. But the convergence is real, and it has a technical cause. The cause is worth understanding, because it tells you which fixes can work and which can't.
Why it happens
A generative model is trained on a large corpus. Roughly: the visible web, plus the popular design systems and component libraries of the last decade. When you ask it for a screen, it produces the most probable one given that training. The most probable screen is the average of what it has seen.
The model doesn't have a style. It has a center of gravity.
You can prompt your way out of the average, locally. Ask for "brutalist, monochrome, dense" and screen one will comply. The problem shows up over time. A prompt is a one-shot correction applied to a system that keeps pulling back toward its mean, and by screen twelve the corners are round again and the gradient is back. I've watched people fight this screen by screen. It works, and it doesn't scale.
Surface tweaks don't do much either. Changing the primary color or swapping the font produces variation, not identity. A visual identity is a few hundred decisions that hold together across the whole product: how type scales, how spacing breathes, how components relate, what happens at the edges. Randomize three of those decisions and you get the average with a different coat of paint.
So how do you get out? It helps that this isn't the first time the industry has faced the problem.
We've been here before
AI didn't invent the sameness problem. I've been building app platforms since 2011, and in the early years the accusation aimed at no-code was, word for word, the one aimed at AI today: all these apps look the same. The cause was different, a finite set of templates instead of a statistical average, but the effect on screen was identical. Pick template 4, get app 4.
At GoodBarber we decided early on not to treat templates as molds. Our templates are recipes: combinations of parameters that our designers have prepared, so you start from something coherent. From there, everything is parameters: thousands of them, down to the shadow on a list cell or the background of a login field. Combined, the possibilities exceed one billion, a number I've had to compute and defend. But the freedom was only half the job. The other half, the invisible one, was making sure that among that billion combinations, the ugly ones simply weren't possible: every parameter is bounded, framed, designed so that any choice lands on its feet. Two apps that leave from the same starting point can end up with nothing visible in common, and neither of them is broken. That's how we set ourselves apart from the template platforms that all looked alike.
The lesson from that era transfers directly to this one. You don't escape sameness by adding more templates, and you don't escape it by prompting harder, which is the same move with new words. You escape it by turning design into a parameterized system, so that identity comes from the combination a product chooses, not from the mold it started in.
The fix is a design system, not a better prompt
That 2010s work left us with more than parameters. It forced us to define an art direction: a signature, a taste, something that makes an app built on our platform recognizable. And it isn't a monument. Our design teams keep it evolving, absorbing the trends of the moment without ever being a copy of them. Formalized, all of that has a name: a design system. Not a theme gallery, not a style-guide PDF: a grammar the platform enforces on every screen it produces.
Ours has three layers: Foundations (grid, spacing, breakpoints, color, typography), ten Atoms (color, font, shape, shadow, border, and so on, the smallest reusable decisions), and UI Components assembled from those atoms. Typography alone is a system of 18 semantic levels, from display headings down to badges, with fixed rules for scale, line height and line length.
Every one of the thousands of parameters from the previous section resolves against those layers.
With AI, the system's job has shifted. A model doesn't produce incoherent screens; coherent-but-average is exactly what it produces. So the design system is no longer mainly there to keep choices from clashing. It's there to carry that signature and impose it on whatever is doing the generating. In the 2010s it guided our users' hands away from the template look. Today it guides the machine away from the average of the web.
And that's the part I want to insist on, because it inverts how most people frame this: a design system isn't a defense against AI. It's what makes AI usable without regressing to the mean.
Speed and identity
The current debate assumes a trade-off. Either you generate fast and look generic, or you design slowly and look distinct. With a formalized system in place, the trade-off mostly disappears. The system holds the identity. The AI fills in content, assembles flows, handles the tedious middle of the work. Everything it produces comes out wearing your grammar, because the grammar is enforced underneath it.
What the system can't do is decide what the rules should be. Choosing which decisions to lock down, and what "right" looks like for your product, is human judgment. That's a question of taste.
Top comments (0)