Design Creativity Bench: 15 frontier models' UI designs fall 0.12 below humans in originality and 0.27 in cross-brief variation, despite exceeding 90% appropriateness
The work introduces Design Creativity Bench, which uses 80 tasks with two open-ended UI briefs per task sharing a goal but differing in product domain, has 15 frontier models each generate 160 HTML designs, and measures originality via embedding similarity (0.592 across models vs 0.764 for human-model pairs) and creative range (0.581 across models vs 0.902 for humans), alongside appropriateness against per-brief acceptance criteria (above 90% for every model, up to 99.2%, vs 98.0% for humans).