Design Creativity Bench finds compliant but convergent model-generated UI
Researchers evaluated 2,400 desktop UIs from 15 frontier models across 80 tasks and 160 prompts. Outputs usually met the brief but trailed human references in cross-model originality and adaptation across product domains.
Why it matters
A renderable, correct interface does not show that a model explored the available design space. The study turns “appropriate but repetitive” into a testable generative-UI risk.
A closer look
The benchmark measures differences between models on the same prompt, variation when the same goal moves across product domains, and acceptance-criteria coverage. The authors report pooled originality of 0.592 and creative range of 0.581, versus human references of 0.764 and 0.902; every model exceeded 90% appropriateness. Fifteen models each generated 160 standalone HTML pages, scored with a segment-based visual similarity method checked against human labels.
Who this is for
Generative-UI researchers, model evaluators, product teams and design-system teams
Original sources
Revisions & corrections (1)
— The arXiv abstract and full HTML paper were read. Submission date, sample size, metrics, pooled results, evaluation method, stated limitations and competing interests were verified from the primary paper. This is a distinct research event not present in the catalog.