Why do LLM pelicans look like this?

AI-generated images and designs are becoming less unique and more generic.

A new benchmark, the Design Creativity Bench, reveals that even frontier AI models like OpenAI's GPT-6 Astra (scoring 65.5) significantly lag behind human designers (scoring 92.3) in creative UI design tasks. The benchmark evaluates originality, creative range, and appropriateness across 80 open-ended UI tasks adapted to two product domains. Despite advances in generating functional UIs, AI models struggle with novel design decisions. The research highlights a "LLM diversity collapse problem" where AI outputs, like SVGs of pelicans, tend to be repetitive. The benchmark's methodology, leaderboard, and evaluation framework are being published to foster measurable AI creativity.

The "LLM diversity collapse problem" observed in image generation, where prompts for common objects yield highly similar outputs across different models, extends to more complex tasks like UI design. This suggests that while AI excels at functional replication, it struggles with the nuanced originality that defines human creativity. The Design Creativity Bench, by measuring originality, creative range, and appropriateness, provides a framework to identify and potentially mitigate this lack of novel output.

This gap in creative capabilities poses a challenge for industries relying on AI for design innovation. Companies aiming for distinctive brand identities or novel user experiences may find current AI tools insufficient for generating truly original concepts. The benchmark's release could spur development of AI models that better mimic human-level design intuition, shifting the focus from mere functional generation to creative problem-solving.