Human diversity fuels collective creativity that large language models cannot simulate or sustain

This preregistered metaphor-writing experiment compared native English (L1) and non-native English (L2) writers working without AI, with AI-generated ideas, or with AI refinement. L2 writers contributed more collective diversity, especially when ideating in their native language. AI ideation compressed diversity and erased the detectable L2 advantage, while AI refinement preserved both. Simulated writer pools built from participants’ real backgrounds, three model families, native-language prompts, and higher sampling temperatures scored below every human pool; pushing diversity further produced degenerate text. AI ideation nevertheless improved individual ratings, exposing a tension between private benefit and collective creativity.