You can spot it in about three seconds. A product shows up in your feed. Before you've read a single word of copy, you already know how it was built: Inter font, a purple-to-blue gradient in the hero, three rounded cards each with a little icon and two lines of text, a headline that says "Build the future of work" and means nothing. It's not ugly. That's the strange part — it's competent. As designer Kasia Gancarz put it, "it just doesn't come from anywhere" [1]. There's even an origin story. In August 2025, Adam Wathan — the creator of Tailwind CSS — posted a half-serious apology on X that got over a million views. He was sorry, he said, for making every button in Tailwind UI bg-indigo-500 five years earlier, which he blamed for turning every AI-generated interface on Earth purple [2]. He wasn't entirely joking. We've started calling it AI slop — and "slop" was literally Macquarie Dictionary's 2025 word of the year [1]. But I want to make a more precise case than "AI is bad at design," because that's not quite true and it's easy to disprove. The real problem is more interesting, and the fix is more empowering than "wait for a better model." Let's get into it. It's not bad. It's average. Here's the reframe that changes everything: AI doesn't produce bad UI. It produces the statistical average of every interface it was trained on. A language model predicts the most likely next token from patterns in its training data. So when you ask it to "make it look modern," it doesn't reason about your brand, your users, or what would make you stand out. It reaches for the safe, universal, offend-no-one choice that dominates its training set — and returns it. Superdesign named this distributional convergence: the model reverts to the statistical center, and the center of "modern web UI" happens to be a very specific, very repeated look [3]. This isn't hand-waving. A 2024 study in Nature documented the underlying mechanism formally — models trained on generated data undergo "model collapse," losing information about the tails of the distribution and converging to substantially reduced variance [4]. In plain terms: the output trends toward the mean and away from anything distinctive. And it compounds — as more AI-generated design gets fed back into the training pool, each cycle makes the median more confident and more homogeneous. The slop gets more polished, which is arguably worse, because polished slop is harder to recognize as slop [1]. So the machine isn't failing at design. It's succeeding at producing the average — and design is one of the few disciplines where being average is the failure. The slop starter pack The tell isn't one thing; it's a constellation, and the community has catalogued it precisely enough that you can use it as a checklist. If your AI-built UI lands even one or two of these, it reads as machine-made [1][3][5]: The purple → blue (or purple → cyan) gradient. The signature move. "Vibecode purple." Inter or Roboto, every time. A centered hero with one CTA, floating in space. A row of three (or six) identical rounded cards, each with an icon, a heading, two lines of text. Glassmorphism with a neon glow. Gradient text slapped on a big number "for impact." A bounce or elastic easing on every hover. Nested cards inside cards. The same corner radius on absolutely everything. Default shadcn-gray and Tailwind-blue, untouched. None of these is wrong. Each is a reasonable, safe choice. That's exactly why they're the average — and why landing a pile of them at once signals "no human decided this." What good design does that AI doesn't Now the contrast, because the point isn't "avoid these patterns," it's understanding what they're missing. Look at the interfaces people actually admire, and you'll notice they're defined by decisions, not defaults. Linear is famous for restraint — a tight, opinionated palette, deliberate density, motion that's fast and purposeful rather than bouncy. Nothing about it is the "average dashboard," because every choice was made against the average. Notion uses color semantically, not decoratively — a color means something (this is a warning, this is a category, this is a state), rather than being sprinkled on for vibes [6]. That's a decision the average can't make, because the average doesn't know what your colors are for. Stripe built an identity out of craft most tools skip entirely — custom illustration, precise gradients used with intent, typography that carries a voice. It doesn't look like a template because it started from a point of view. The common thread: good design comes from somewhere. Someone decided what this product should feel like, who it's for, and what it should emphasize — and then made a thousand small choices in service of that. The AI has access to none of that, because none of it exists in the prompt "build a pricing page." As one anti-slop guide put it bluntly: "The root cause of slop is no decision — defaulting to the statistically safe look" [5]. Which brings us to the thing this article is actually about. The real fix isn't a better prompt. It's your taste. Here's the uncomfortable, empowering truth: this problem is structural, and it will not be fixed by a cleverer one-shot prompt or a smarter model [3]. The model knows the syntax of good design — it can produce clean, valid, plausible interfaces all day. What it can't do is know which decision your project should make. That decision is taste, and taste is the one thing you have that the machine doesn't. That's not a consolation prize. It's the whole job now. AI made competent-average design free, which means the scarce, valuable thing is no longer the ability to produce a clean interface — it's the ability to decide what this specific thing should look like, for these specific humans, and why. Distinctive design has quietly become a moat, precisely because everyone else is shipping the default [7]. So the fixes below aren't tricks to make the AI more creative. They're ways to inject your creativity into a machine that has none. The AI is the hands; you have to be the eyes. How to actually fix it In rough order of impact: 1. Write down your design system and make the tool follow it. This is the single highest-leverage fix. Create a DESIGN.md (or a tokens file) that specifies your colors, type scale, spacing, and radius, and feed it to the agent as a constraint it must obey [6][8]. Without this, the AI invents spacing on component one, invents different spacing on component two, and by component five your UI is a patchwork no amount of "make it look professional" can rescue [8]. With it, every component inherits your decisions instead of the average. Own your tokens as CSS variables so the direction propagates automatically. 2. Commit to ONE aesthetic direction — and "clean and modern" isn't one. "Clean and modern" is not a direction; it's the slop default in a trench coat [5]. Before you style anything, pick an actual point of view — brutalist, editorial, warm and playful, dense and technical — and lock its tokens. A committed direction, held consistently, is what makes a thing look like it came from somewhere. 3. Use color semantically, not decoratively. Steal Notion's discipline: a color should mean something [6]. Primary action, warning, success, category. When color carries function instead of just vibe, the interface stops looking like a gradient was applied for fun and starts looking like someone thought about it. 4. Separate the jobs — don't ask one prompt to do taste, exploration, and code at once. "Build a pricing page" asks the model to make design decisions, implement them in production code, and fit your system, all in one shot [8]. Split it: decide the creative direction in plain text first (that's your job), then have the agent implement the direction you chose. Taste and typing are different tasks; stop fusing them. 5. Prompt for divergence, not the first answer. Two cheap moves that push the model off the average: ask for three genuinely different directions instead of one (it'll explore more of the space instead of settling on the first local maximum — which is always purple) [2]. And assign a persona — "you're a senior designer with a background in print and editorial" shifts the probability distribution it samples from [2]. Neither gives the model taste, but both stop it from defaulting. 6. Keep an anti-slop banned list. Take the slop starter pack from earlier and treat each item as a hard fail in review [5]. Purple-cyan gradient? Reject. Gradient text on a metric? Reject. Bounce easing everywhere? Reject. A short banned list, applied every time, catches the machine-made tells before they ship. The bottom line AI didn't make design worse. It made average design free — and in doing so, it quietly raised the value of everything average can't do. The purple gradient isn't a bug in the model. It's what you get when nobody decides. The interfaces that stand out — Linear, Notion, Stripe, and whatever you're about to build — stand out because a human brought a point of view the machine could never have, and then used the tool to execute it. So let the AI do what it's good at: turning your decisions into clean, working components, fast. But the deciding — the taste, the intent, the willingness to be specific instead of safe — that part was always yours, and now it's the part that matters most. The model gives you the average. You're the reason it doesn't have to stay there. What's the telltale sign you use to clock an AI-built UI in three seconds? I'll start: gradient text on a big number, and that exact shade of Tailwind blue. Add yours below — let's build the definitive slop banned-list in the comments. References Gancarz, K. "Spot the Slop: A UI Designer's Guide to Fixing AI Defaults." Mania Design, July 2026. (The 3-second recognizability of AI UI; "it doesn't come from anywhere"; "slop" as Macquarie Dictionary 2025 word of the year; semantic color.) "Why Your AI Keeps Building the Same Purple Gradient Website." prg.sh, October 2025. (Adam Wathan's Aug 2025 Tailwind bg-indigo-500 apology, 1M+ views; persona prompting; requesting multiple directions.) Superdesign. "Why AI Design Looks Generic (2026)." June 2026. (Distributional convergence; structural default, not a capability gap; fix is process and context, not a cleverer prompt.) Shumailov, I. et al. "AI models collapse when trained on recursively generated data." Nature, 2024. (Model collapse; loss of distribution tails; convergence to reduced variance.) Vibe Code Kit. "AI Slop Design: Why AI-Generated UI Looks Generic (Fix Guide 2026)." June 2026. ("The root cause of slop is no decision"; the anti-slop banned list; commit to one direction.) Braingrid. "Design Systems for AI Coding: Stop Getting Purple Gradients." December 2025. (DESIGN.md / design-system files, CSS variables, tokens the agent follows.) eChai. "How do I stop my AI-generated UI from looking generic?" (Distinctive design as a moat; separate taste, exploration, and code; DESIGN.md.) Lavaee, A. "Why My AI-Generated UI Looked Generic (and How I Fixed It)." (The patchwork-by-component-five problem; asking one prompt to decide, implement, and integrate at once; feeding an extracted design definition back to the agent.) Note: figures and phenomena here are drawn from design-practitioner writing and one peer-reviewed study across 2024–2026; the named products (Linear, Notion, Stripe) are cited as widely-recognized examples of intentional design, not as endorsements or affiliations. Follow the links for each source's full argument.