All notes

The AI design monoculture is measurable now

For about two years I have been saying that AI-generated interfaces converge on one look — same cards, same gradients, same safe defaults. That line used to be a vibe and I had to defend it as one. Not anymore. The new Design Creativity Bench from Apple researchers put numbers on it, and they are worse than I expected: same-prompt UI designs from different models overlap at 0.592 originality in the Design Creativity Bench paper, against 0.764 for human-model pairs. I wrote earlier this month that the median AI product is getting worse — this is the measurement underneath that claim, and it says the monoculture is structural, not a coincidence.

What the numbers say

The same benchmark, Design Creativity Bench scores three things. Originality — how different two designs are when different models get the same prompt — came in at 0.592. Creative range — how much a model changes its design when the same UI goal moves to a different product domain — came in at 0.581 against 0.902 for human designs. Both fall dramatically below the human baseline.

Here is the part that stings: appropriateness was fine. Every model tested cleared 90% of the brief's acceptance criteria, and the best hit 99.2%, slightly above the human baseline. So the average AI design is competent and repetitive simultaneously. It meets the spec and looks like the other four tabs you have open. That is exactly the trap: it passes your review because nothing is missing, and it loses in public because nothing is memorable.

The preference vote is the real story

The team at HumanSignal ran a separate piece worth reading: they judged designs from the major models blind and found people preferred the models that broke from the monoculture — not the ones with the highest raw eval scores. When you remove vendor names and put two real interfaces side by side, the average one reads as wallpaper. The one with a point of view gets remembered.

That matches what I keep seeing. Nobody bookmarks a product because the padding was correct. The models most heavily aligned for "polished" output are also the ones most aligned for "most probable" output, and those are two names for the same failure.

How I get out of the monoculture

The fixes are cheap and they go in before generation, not after. Three moves I use:

  1. Write an aesthetic posture down. Pick a stance the model must hold — brutal, editorial, slate-and-amber, whatever — and put it in a file the model reads every session. Vague "make it distinctive" gets you the same gradient. A named stance does not.
  2. Ban the defaults explicitly. The anti-monoculture moves are negative constraints: no Inter, no purple-to-blue, no 24px-of-everything cards. The model follows bans far better than aspirations.
  3. Generate from your tokens, not the corpus. Colors, radii, and motion referenced from your design tokens leave the model no room to fall back to its training average.

None of this makes the model creative. It makes your defaults the only distribution the model can sample from — which turns out to be enough. The human-model pairs in the same benchmark measurements hit 0.764 originality, and those humans were not magicians. They were just holding constraints the models could fail against.

Is originality worth chasing if the output is already good?

Yes — because "good" is the entry fee, not the win. A 2026 solo product competes in a feed full of apps that are all competent, so the tiebreaker is recognition in one glance. The models that broke from the pack won the votes. Yours can too.

Frequently asked questions

  • The tendency of LLM-generated interfaces to converge on the same look — rounded cards, a Tailwind SaaS layout, one safe accent color — regardless of who asked or why. A 2026 benchmark measured it directly: same-prompt designs from different models overlapped far more than human designs ever do.

About the author

mosh

mosh is a product designer for growth, working with design thinking and ever-improving design systems. What matters: fixing conversion, whether in B2B dashboards or direct-consumer apps.