01 / DEFINE THE KIND OF CREATIVITYA good idea can be novel, useful, or both—and those are different tests.
Creativity is often assessed through originality and usefulness. A study may ask independent readers to rate a story, count unusual answers, or judge whether a proposal solves a real brief. These measures capture different qualities. A higher score on one task does not establish that a person or a system is more creative in every setting.
NoveltyDoes the work bring a meaningfully new direction for this task and audience?
UsefulnessDoes it meet the brief, constraints, and practical purpose?
Individual qualityHow strong is one person’s finished work when judged on its own?
Collective diversityHow much do the final ideas differ across a group? Similarity can rise even when each idea is rated well.
When AI is part of a creative workflow, keep these outcomes separate. A polished first draft can save effort while still narrowing the set of directions a team considers.
In one story-writing study, individual ratings rose while stories grew more alike.
Doshi and Hauser randomly assigned 293 writers to produce an eight-sentence story with no AI idea, one AI idea, or access to several ideas from GPT-4. A separate group of 600 readers evaluated the stories. The AI-idea conditions received higher ratings for creativity, writing quality, and enjoyment, with larger gains among writers who scored lower on an initial creativity measure.
The same experiment found that AI-assisted stories were more similar to one another than stories written without AI ideas. That is a group-level result from one short, English-language writing task with online participants. It does not establish that every AI tool or creative field produces the same trade-off.
Evaluate the whole collection, not just its highest-rated entry. A team can improve its average draft and still lose some of the unusual directions that make a portfolio valuable.
Access alone did not determine who benefited.
A 2025 field experiment randomly assigned 250 employees at a technology consulting firm in China to use ChatGPT during a work week or continue without it. Supervisors and external reviewers rated the employees’ creative work. The results depended on metacognitive strategies: people who planned, monitored, and adjusted their approach were more likely to show creativity gains with LLM support.
This is evidence from one firm and a short workplace study, not a guarantee for every role. It does point to a practical distinction: giving people access to a generator is different from helping them direct, question, and revise its suggestions. For a related guide to those skills, see AI literacy at work ↗.
The average is promising, but the evidence is still uneven.
A broad 2024 meta-analysis of 106 human–AI experiments found that combined systems generally performed better than people alone but, on average, fell short of the better solo performer. Most measured tasks involved choosing among predefined options; creation tasks were less common, so its overall result does not settle how generative AI affects creativity.
A 2025 preprint focused specifically on creative tasks and synthesized 28 studies with 8,214 participants. It estimated a small average gain for human–GenAI creative performance. Its analysis of idea diversity, however, relied on only six comparisons from four studies, with substantial variation between them. Treat that result as an early signal for further research, not a universal prediction about creative work.
For a deeper look at the distinction between helping a person and improving the combined system, continue to how to evaluate human–AI collaboration ↗.
