待翻译：AI-generated websites converge on each other more than human designs do

AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。

Machine-made websites converge on each other A measurement, not an opinion 945 sites · 5 tools · Aug 2026 Machine-made websites converge on each other I gave five AI tools an identical brief, then measured 945 award-winning websites on fourteen dimensions to see where the machine output landed. It landed in the middle — and, more tellingly, it landed together. The findingPairwise distance in normalised feature space 4.45 Median distance between two random award-winning sites 1.91–3.53 Distance between any two AI outputs Three different companies. Three different models. Three different technology stacks. Given the same brief, they produced pages that sit closer to one another than two human-designed award-winners typically sit. That is what convergence looks like when you measure it rather than assert it. The tools are not copying each other. They are each independently walking to the middle of the same room. The map945 measured sites, projected to two axes Every dot is a real website pulled from Awwwards, Httpster, One Page Love, Land-book, recent.design and SaaS Landing Page, loaded in a real browser and measured after its fonts finished loading. The three marked points are AI output. ← fewer type sizes, smaller palette, stricter grid more sizes, richer palette, looser grid → 945 award-winning sites AI output Principal components of 14 measured features. Horizontal axis is dominated by type-size count, palette size and grid adherence; vertical by spacing discipline and padding variety. The AI points do not sit at the edge of the human cloud. They sit in the thick of it. TypicalityDistance from the centre of human practice If machine output were obviously bad, it would sit far from the centre. It doesn't. Two of the three are more typical of award-winning web design than the median award-winner. Distance from the centroid of 945 sites (lower = more typical) Subjectz-distance Claude Design2.16 Replit2.34 Median human site3.01 Codex3.30 90th percentile human site4.31 Most unusual human site24.65 The machine output is not bad. It is central. Which is exactly what a process that predicts the most likely next thing should produce. The briefIdentical text, no style words, first output only Every tool received this and nothing else. No "modern", no "clean", no "premium" — those words steer every model to the same place and would have manufactured the result. Build a one-page website for Halvard & Sons, a two-person workshop in Sheffield that makes solid brass mechanical pencils. One product, the No. 4, £180, made to order, six-week wait. The page needs to cover: what the pencil is, how it's made, why it costs what it costs, the six-week wait, and how to order. There's a 15-year-old daughter who does the packaging illustrations. Mention that somewhere. Real copy, not lorem ipsum. Desktop layout. The brief was chosen so that nothing in it could be solved from a template. The six-week wait and the daughter are the parts a system has to make a judgement about. The outputsSame brief, three of the five tools Nobody used Inter. Nobody made a purple gradient. Nobody shipped the rounded-card grid. The 2026 checklist for spotting AI design is already out of date. Instead, four of five independently produced a warm brass accent on a warm ground with a serif display face and uppercase letterspaced labels — two of them landing on the same typeface, Fraunces, with no coordination. Lovablecream · serif display · brass Figma Makewarm black · Fraunces · brass Codexneutral · system sans · photography The tellsWhere machine output actually differs Two measurements separated the AI output from human practice, and neither is about taste. Percentile within the 945 human sites MeasureHuman medianAI output Largest corner radius50px0px · 0th pct Grid adherence0.620.78–0.92 · 70–82nd pct Distinct type sizes810–13 Gradient elements00–3 Perfectly sharp corners, and a grid it refuses to break. The most expressive cluster of human sites has the lowest grid adherence in the whole dataset — those designers deliberately break their own columns. The machine never does. Everything else — type scale, palette size, spacing rhythm, density — is indistinguishable. MethodWhat was measured and how 945 sites harvested from six public galleries by following their outbound links; no screenshots taken, no paywalled sources used. Each loaded in headless Chromium at 1440×900, measured only after document.fonts.ready — measuring earlier returns fallback-font metrics and quietly corrupts every type number. Fourteen features per site: type-size distribution weighted by how much text uses each size, colour weighted by painted area, section-padding discipline, radius, grid adherence, above-fold density, media counts. Features winsorised at the 1st and 99th percentile and log-compressed where skewed. Before this correction, twelve of fourteen features looked bimodal; afterwards only five did. The difference was a handful of sites using border-radius:9999999px. Control: clusters were checked against the source gallery. Purity ranged 0.29–0.35, so the clusters are not simply reproducing which gallery a site came from. Had they, this would measure curation rather than design. Limitations. Three of five AI outputs could be measured — one produced only screenshots, one only pasted source. One brief, one industry, one viewport. The convergence result rests on three pairwise distances, not thirty. AsideFound while trying to build the outputs Replit's export does not build outside Replit. It ships catalog: and workspace:* version specifiers that only resolve inside their monorepo, imports a private package that isn't on npm and isn't used by any source file, extends a tsconfig two directories above the zip, and requires a Replit-only Vite plugin. It also imports two hero images it generated — hero-no4-pencil.jpg and lathe-workshop.jpg — and ships neither. It also shipped 55 UI components for a static one-page site, including a carousel, a calendar and a one-time-password input. None are used.