AI slop is what the average looks like

You know the page. A purple-to-blue gradient behind a headline that ends in reimagined. Three cards in a row, each with a rounded icon, a two-word title and a sentence about being seamless. Frosted glass. A soft glow. A button that says Get started, and another one that also says Get started.

People call it AI slop. I think the name is fair, but the diagnosis usually isn't. The common story is that the models have bad taste. I don't think that's true. I think the models have the average taste of everything they've seen, and the average is exactly what this looks like.

Why everything converges

A model that generates interfaces has learned from an enormous number of existing ones. When you ask it for a landing page for a project management tool with no other direction, it gives you the most likely landing page for a project management tool. The most likely page is the one that resembles the most pages. It's the centre of the distribution.

The centre has a look. A few years of startup templates, UI kits and modern SaaS tutorials have taught it that modern means gradients, rounded corners, glass and an icon above every heading. So that's what comes back, and because everyone is prompting from the same centre, everyone gets the same page.

Copy has the same problem, and it's worse. The most likely way to describe any product is the way that could describe every product. Streamline your workflow. Built for modern teams. The all-in-one platform. These phrases survive because they're never wrong, and they're never wrong because they don't say anything.

So slop isn't a failure of capability. It's what you get when nobody states a point of view. The model fills the gap with the mean.

Slop is a design problem

This matters because it tells you where the fix lives. A better model, on its own, won't solve it. A better model will produce a more polished average.

The fix is the same one designers have always used against templates: a stated point of view, applied consistently. A brand that decides it will never use gradients. A voice guide that bans seamless. A rule that every button names the outcome it produces. These decisions are what make a product look like itself instead of like the category.

What's changed is who needs to read those decisions. It used to be enough for the design team to hold them in their heads and enforce them in review. Now the agent producing most of the first drafts needs them too, in a form it can act on, and you need a way to check whether it did.

The tells are measurable

Here's the part that surprised me when I started building checks for it: most of the tells of generated design are not matters of taste at all. They're patterns, and patterns can be detected in code.

Gradient-filled text is a CSS property. A purple gradient is a set of colour values. Frosted glass is a backdrop blur. A coloured glow is a box shadow with a large blur radius and a saturated colour. A row of identical icon cards is a DOM structure you can recognise. Get started is a string.

When I wrote the slop rule packs for Glitterball, 23 of the 28 rules ended up being measured in code, with no model involved. They cover three groups:

• Visual: gradient text, stock purple gradients, glass, neon glow, emoji used as icons, sparkle icons, blur blobs, rows of identical feature cards, every container heavily rounded, long centred paragraphs, too many typefaces.
• Copy: a list of buzzwords, stock headline phrases, overuse of em dashes, it's not X, it's Y constructions, three-word taglines, generic button labels and claims that can't be checked.
• Unfinished: lorem-style filler, stand-in names and stock images, links that go nowhere and the same sentence repeated in two sections.

Only a handful of rules really need judgement: whether the copy is specific to this product, whether it reads as one voice, whether the page looks like a template overall. Those go to a model, and only those.

A rule is a point of view you can enforce

Each of these rules is a small, opinionated decision. You might disagree with some of them. You might love glass. That's fine, and it's the point: the rules are switches, and you add your own in plain words. Don't use the words synergy or leverage becomes a check that runs against every line of copy on the page.

What you can't do any more is leave the decision unmade. An unmade decision gets made for you, by the average.

I find this a useful way to talk to teams about AI-generated design. The question isn't is the model good enough? It's have we told it anything? Most teams haven't. They have a brand guide written for people, a component library, and a lot of taste that lives in a few senior heads. None of that reaches the agent.

What this site checks

This site is built to pass its own slop packs. There are no gradients in the type, no glass, no glow and no row of three icon cards. The buttons say what they do. Before each change ships, I run the same checks against the page that Glitterball runs against yours.

That's not because I think every page should look like this one. It's because a rule set that the author's own work can't pass isn't a standard. It's a wish list.

In the next piece I'll look at how AI labs evaluate their models, and which of their habits are worth stealing for design.