The Emperor Pelican Has No Clothes

A striking demonstration shows that when asked to "Generate an SVG of a pelican riding a bicycle," many large models produce near-identical, homogenized images instead of varied creative outputs. Side-by-side comparisons of model runs yield repetitive pelicans, while a richer set of plausible alternatives - royal parade, monsoon, polar night, alpine race, desert, fog, neon alley, wedding, market, Victorian promenade, storm and Moon - highlights how much visual diversity is lost. That pattern is described as an "LLM diversity collapse" and positioned as a broader creativity failure rather than a mere rendering quirk.

To quantify this shortfall, researchers developed Design Creativity Bench, evaluating 16 frontier models on 80 open-ended UI design tasks across two product domains with human designer references. The benchmark scores models on originality (distinctiveness across outputs), creative range (adaptation across domains), and appropriateness (meeting functional brief). Results reveal a large gap: the human reference scores 92.3, while top models lag - GPT-6 Astra 65.5, GPT-6.1 Sol 63.7, Grok ~62.2, and Gemini Flash 60.6. The full leaderboard, methodology, datasets and reproducible evaluation framework are published, demonstrating that current models can produce functional interfaces but still struggle to make genuinely creative design decisions.