Design Creativity Bench

As LLM adoption in UI design accelerates, we are entering an era of design monoculture - what the internet increasingly calls “AI slop.” Our data backs this up: The overall score, which weights both diversity and appropriateness, reflects the exact same story. GPT-6 Astra tops the models at 62.9, but the human reference sits far ahead at 84.6.

When we asked different models to build a screen for the exact same brief, they clustered around a highly repetitive default, with a median score of 0.591 on distinctiveness. Introduce a human to the mix, and distinctiveness jumps to 0.767. Worse, models don't adapt to context. Ask a model to design for a clinic versus a pharmacy, and its designs barely change (0.581 median distinctiveness). A human designer intuitively shifts their approach to match the product, resulting in a distinctiveness score of 0.902. The models have the capacity for range, but without steering, their default state is a sea of sameness.

The uniformity of AI design isn’t a sign of failure. In fact, the models we tested were remarkably good at doing precisely what they were told. Every one of them satisfied at least 90% of a brief’s requirements, and the best of them (Claude Opus 5.5 and GPT-6 Astra) outperform human designers at following instructions, with 99.1% appropriateness.

When models do falter, their mistakes tend to be small and specific. In 31% of cases, the failure involved leaving out a control tied to a specific part of the brief. In 27% of cases, the models generated sample data that didn’t quite add up. We also noticed that with each new generation, models tend to become slightly better at following instructions, and slightly less capable of producing a variety of designs. The latest GPT-Sol, Claude Opus, and Grok models, for instance, all outdo their predecessors in building what they were asked for, but they have the same or less creative range. The question now is not whether LLMs can design an interface that matches the brief, but whether they can do it with human-level spark - either through human direction or with increased creative capability of their own.