Forty concepts, three ideas

The test

We took two real competitions with public outcomes. The 2024 Helsinki Architecture and Design Museum competition has 615 entries we could rebuild, with the jury’s tiers. The eVolo Skyscraper Competition has 253 awarded entries from 2013 to 2025. For each brief we made 40 concepts six ways and set them against 40 real entries drawn at random.

The six ways are plain prompting, Verbalized Sampling (the model gives several answers with their probabilities), a random concrete seed for each idea, 40 invented people’s frameworks with one idea each, a dial that asks for unusualness from level 1 to 5, and an engine that keeps the 40 of 120 candidates farthest from the real entries.

Distance is measured in an embedding of each concept’s one-line summary and core idea, against how far real entrants sit from each other. Zero is a typical entrant, and the far band is the real field’s top tenth. How many different ideas a set holds is counted with the Vendi score. Both are screens for difference. Neither says an idea is good.

Plain prompting collapses

Forty plain concepts hold about 2.7 distinct ideas on Helsinki and 3.8 on eVolo. Forty real entries hold 10.8 and 16.2. Plain concepts also sit closer to the typical entrant than real entrants do.

A different seed per idea restores the spread

Seeded arms hold 7 to 14 distinct ideas in 40. Forty invented people’s frameworks give 9.2 on Helsinki, close to the real field’s 10.8, and 92% of their concepts land in the real field’s far band. On eVolo the engine gives 13.9 and the frameworks 10.4. Verbalized Sampling also spreads, but its tails converge. Most of its 40 Helsinki concepts belong to one family, a seasonal ice or floating barge.

Random seeds spread ideas about as well as a real person’s life writing did in an earlier test. So the spread comes from the seed, not from the person, and a person’s framework has to earn its place some other way.

The dial moves ideas, but not far

Asking for unusualness from level 1 to 5 tracks measured distance from the real entries on Helsinki (Spearman 0.89 over 40 concepts, from z −1.70 at level 1 to +0.57 at level 5). Even level 5 stays inside the field’s normal range. On eVolo the dial tracks less well (0.68, with levels 2 to 4 flat), and on New Yorker captions it is weak (0.22).

What the judge says, and why to be careful

No AI arm beats a real upper-tier entry by our judge, a Claude model that reads each concept and a real entry restated in the same words, in both orders. On Helsinki plain concepts win 25% of the time, the engine 22%, random seeds 18%, the dial 16%, Verbalized Sampling 14% and the frameworks 2%. On eVolo the rates run from 0% to 31% against an honourable mention and 0% to 24% against a podium entry. Across AI concepts, distance and preference move in opposite directions (rho −0.21 on Helsinki, −0.28 on eVolo). The next paper takes that up.

The judge is a weak instrument. It sides with the real jury 68% of the time on top against bottom tier Helsinki entries, and 59% on eVolo podium against honourable mention, an interval that includes 50%. Asked which of two concepts was written by the human team, it picks the real entry 98% of the time on Helsinki and every time on eVolo. So “prefers the real entry” and “spots the AI” cannot be told apart here.

What it means

The machine can place concepts almost anywhere on a real field’s curve on request, and 40 of its concepts can be as varied as 40 real entries. That is the part we can show. Whether people like the far concepts is another question, and only a blind round with architects can answer it. So far this is two briefs, one judge and no architects.

Next: Far is not new, and far costs fit

All papers