Human direction drives creativity with large language models

Large language models (LLMs) can generate ideas, stories, and other creative content with seemingly minimal human input. Yet whether creative work can now be delegated to LLMs—or whether human direction remains crucial—is currently unclear, partly because we lack tools to measure how people collaborate with them. Here we introduce Creative Direction (CD), a construct scored automatically from human–LLM chat transcripts that captures the extent to which people direct rather than delegate creative work by contributing their own ideas and critically refining model suggestions. We studied CD across two controlled experiments using ChatGPT (N = 234; N = 205), a preregistered replication with Google Gemini (N = 157), and 6,290 naturalistic chat interactions. Across all samples, people who exercised more CD during brainstorming and writing tasks consistently produced more original content with LLMs (meta-analytic r = .28, 95% CI [.25, .32]), an effect that held beyond simple interaction frequency. People who were already more creative without LLM assistance performed better on creative tasks with LLMs, which was mediated by their tendency to direct more. CD also countered the so-called “homogenization effect”—the tendency for LLMs to generate the same kinds of ideas across users: higher direction reduced semantic similarity in the group, making people’s ideas more distinct. We therefore show that creative outcomes with LLMs are shaped by people’s tendency to direct or delegate the creative process.