Far is not new, and far costs fit

Far costs fit

In both architecture competitions we tested, the further a concept sits from real entries, the less an AI judge prefers it. Across AI concepts the rank correlation between distance and preference is −0.21 on Helsinki and −0.28 on eVolo.

Framework’s Ideas page shows the same trade on its own scale. It makes concepts in five bands, from a plain answer to the farthest, and a separate judge scores each one from 0 to 100 for how unusual it is and how well it fits the brief. Over 90 concepts for the Helsinki brief, unusualness rises from 25 to 66 across the bands while fit falls from 78 to 53. The farthest concepts look about 25 points less buildable than plain ones. The bands order concepts as intended (Spearman 0.77), so the dial works. The price of turning it up is fit.

Design research found the same trade

This is not a quirk of our judge. Studies of design by analogy, where designers draw inspiration from examples near to or far from their problem, report the same pattern. In a project with 105 student designers, novelty rose as the inspiring patent grew more distant, and quality rose as it grew closer. In a study with 96 participants, near stimuli improved the feasibility and usefulness of solutions and distant ones improved their uniqueness. Another team found that stimuli can be too far, and then they harm the process. Our result is the machine version of an old finding, which makes us more sure the trade is real. What other researchers found has the sources.

Far from the field is not new to the world

We also asked whether the far concepts are new. An AI checker with web search named the nearest built or published precedent for 60 real Helsinki entries and 120 concepts from our arms, and rated how close each came. We tested it on 20 planted famous buildings. It named the right building 39 times of 40, faithful and disguised, and rated none new.

By its count 5% to 10% of real entries are new, which here means the nearest precedent is only a family resemblance. Plain prompting scored 5%, Verbalized Sampling 0%, random seeds 5%, the dial 15% and our far-tail engine 0%, with 7 of its 20 concepts close to a known building. Plain Claude named the Oslo Opera House for 12 of 20 concepts. The frameworks scored 40%, but those were budget rules and programme ideas with no building form to match.

So distance from the field is not newness against the built world. The checker is a model too. On a retest it gave the same new-or-not call on 25 of 28 concepts and the same precedent on only 13 of 30.

More rounds keep ideas far, not better

Revising concepts with real outside material, a research abstract each round, kept far ideas far. After four rounds 109 of 138 variants were still far from the field, against 40 of 141 in a closed loop. A random fragment did nearly as well (104 of 141).

But the AI judge liked the concept less after four rounds in every arm. Against a real upper-tier entry it fell from 21% at round 0 to 2% to 4% at round 4. Distance and preference are unrelated across rounds (rho 0.00). Selecting on preference first stops the slide and does not make the concept climb (21% at round 0, 8% at round 4, with an interval from 0 to 23%).

What it means

Far ideas cost fit, and distance alone is not novelty. That is why the product is the chooser and not a bigger generator. Something has to find, for one person, which far ideas are worth the cost. Whether people value far ideas anyway, and whether development makes them better, is what the blind round with architects is for.

Next: A selector that learns one person

All papers