Why every AI-generated UI looks the same
Purple-to-blue gradient cards, a centered icon grid, low-contrast dark mode. The tell is hardcoded hex drifting away from a token file nobody is checking against.
Every AI-generated UI I see lately looks roughly the same: a particular sans-serif, purple-to-blue gradient cards, a centered icon grid, a low-contrast dark mode. One AI lab has a name for this: distributional convergence. Without explicit steering, a generative model samples from the safest, highest-probability region of its training distribution — the average of everything it has ever seen.
I ran into this building hop, my flights side project, and while building this site. You prompt, you get something coherent, and then you notice it looks like everything else on the internet right now.
The scale of it is clearer with numbers. An analysis of 1,590 Show HN vibe-coded submissions found 22 percent showed heavy slop, four or more recognizable patterns stacked together, 32 percent mild slop, 46 percent clean. More than half of publicly shipped projects lean on the same handful of defaults. No single fix solves it. The countermeasures that actually move the needle share one property: they're structural, not a better prompt.
The most reliable lever is grounding generation in a real component library instead of letting the model synthesize freely. In hop I run a /design-sync step that pulls a set of locked primitives into the generation context before anything gets built, so the model has no route into generic territory to begin with. The same approach works with a known library like shadcn/ui, or a Figma- or Storybook-ingested one. The constraint has to be external and enforced, not requested.
Token discipline helps almost as much. Hand a model raw hex values and it hardcodes them, and drift follows within a few generations. The fix is a tiered vocabulary: raw color primitives stay hidden from the model entirely, semantic tokens become its actual working vocabulary, component tokens sit on top of those. Pair that with deterministic linting — not another model — to catch drift the moment it happens rather than three components later.
the model that sees hex codes hardcodes them; the model that only ever sees semantic tokens can't drift, and a lint pass catches it if it tries.
The third countermeasure is forcing an aesthetic-direction decision before generation starts, grounded in the brief's own subject matter rather than a default. Brainstorm the full system first, then review it against the brief specifically for what makes it distinct. Restraint is explicit guidance here, not an afterthought: concentrate the boldness in one place and keep the rest disciplined.
On AI-judged design quality, I treat it as an open problem rather than a shortcut. A model's critique correlates weakly with human judgment — Spearman 0.194, against 0.556 for human critiques. Even trained designers often disagree with each other on which of two AI designs is better; Krippendorff's alpha lands at 0.248 on that question. What actually produces stronger results is fine-tuning on rich human feedback, sketches and direct revisions, where designers agreed 76 percent of the time, well ahead of ranking feedback or automated judge loops. The fixes are partial, annoying to maintain, and the evaluation problem stays genuinely unsolved.
The leverage is in constraints and process: grounding, token discipline, explicit restraint, skepticism toward automated evaluation. Everything else is hoping.