The Emperor Pelican Has No Clothes

Researchers have identified a "diversity collapse" problem in large language models, demonstrating that AI systems generate remarkably similar outputs when asked to create designs—such as pelicans riding bicycles—despite the potential for much greater variety. A new Design Creativity Bench benchmark evaluated 16 frontier AI models on UI design tasks across three dimensions (originality, creative range, and appropriateness), finding that the best-performing model scored only 65.5 out of 92.3, significantly underperforming human designers in creative decision-making.