The Website Challenge

Typography was the one variable a human deliberately left open, with an explicit warning against the obvious answer, and all four AI design tools chose the same typeface anyway. One person wrote a design brief once, froze it, and handed the identical frozen document to four different AI design tools with nothing else varied. Across six independent readings of that brief, the display typeface was Archivo every single time. Whether that is genuine agreement on the right answer or four tools pattern-matching to the same training-data reflex is the open question this run cannot settle by itself, which is why the four exhibits are published for readers to judge directly.

What was the Website Challenge?

One person wrote what he wanted from a website, once, in his own words. That document (VISION.md) was frozen before any tool touched it. Four AI design tools were each handed the same frozen brief, the same starting code, the same stack (Astro 7 static, Tailwind 4), the same four page types (home, blog index, blog post, about), and the same content. Nothing varied except the tool. Each tool ran in an isolated worktree, in a fresh, clean-context session that had never read any tool's rules or any other arm's output.

What did every tool choose, independently?

The brief's instruction on type was direct: "I don't want to restrict, I want to be surprised," paired with a named embarrassment to avoid, "an obvious serif font selected by AI." Type was expected to be one of the places the four tools would diverge most. It was not. All four tools chose Archivo for display type. Arms 1 and 2 independently chose the identical pair, Archivo and Faustina, with no contact between them. Inside arm 4 alone, three further independent voices each proposed Archivo without seeing one another's answers. Six independent readings of the same brief, one display typeface, every time.

That convergence is the most interesting thing this run produced, precisely because it cuts both ways. Read generously, it looks like four independent systems finding the same correct answer to a well-specified problem. Read skeptically, it looks like four tools built on similar training data reaching for the same safe, non-obvious choice: "surprise me" answered with consensus instead of variation. The run itself cannot distinguish between those two readings. That is exactly why the four exhibits are left live: so a reader can look at the actual outputs and decide which explanation the evidence supports.

The four exhibits

Each card is a full, independently built website from the same frozen brief. All four exhibits are frozen and noindex: they will not change, and they should not be relied on to appear in search results. This hub carries the canonical explanation and ranking of the run.

If the outputs converged, what actually differed?

The divergence between the four tools was in process, not in taste. The cheapest arm (frontend-design) modelled at $1.12; the most expensive (Impeccable) modelled at $7.02, a 6x spread, yet the visual outputs were more alike than that spread would predict. What this run actually publishes is that process divergence, plus the single most valuable defect it uncovered: a colour that, due to a CSS specificity conflict, could structurally never render, a problem only an adversarial review pass was able to find. Full detail, including the per-tool process breakdown, the defect, the stats tables, and the cost modelling method, is on the results page.

Read more

• Methodology: how the run was controlled. Clean-context runners, the intake proxy, parallel execution, and verification.
• Results: the headline finding in full, the process divergence between tools, the key defect, the stats tables, and the cost method.
• Exhibits: the four tools with per-tool context, per-arm stats, and links out to each live site.