Large Language Models Exhibit Homogeneous Creative Outputs

Researchers Emily Wenger and Yoed N. Kenett report in PNAS Nexus that across standard creativity tasks, outputs from many LLMs—including Gemini, GPT, and Llama—are significantly more similar to each other than human responses. While single AI answers can score as creative as humans, population-level homogenization persists; raising model temperature increases variability but quickly produces gibberish, implying LLMs may narrow brainstorming and creative diversity.