Measured by output, AI can already be creative on some narrow tests: in published experiments, chatbots scored higher than the average person at inventing uses for everyday objects, and readers could not reliably tell chatbot poems from poems by famous poets. Measured by process, intent, and independence, the question is open, because no study has shown that a model sets its own goals or knows why a work matters. Researchers also find a cost: stories written with AI ideas were rated better but more alike.
Key takeaways
- A 2023 Scientific Reports study compared 256 people with three chatbots on an uses-for-objects test, and the chatbots scored higher on average while humans still produced most of the best ideas.
- In a 2024 Scientific Reports study, readers identified AI-written poems correctly only 46.6% of the time, below chance.
- A 2024 Science Advances experiment found AI story ideas raised creativity ratings and also made stories more similar to one another.
- Each test measures one slice of creativity, so none settles the broad question.
What "creative" means in these studies
Psychologists usually treat a product as creative when it is new and also useful or fitting for its purpose. The 2024 Science Advances paper by Anil Doshi and Oliver Hauser rates stories on exactly those two dimensions, novelty and usefulness. That definition judges the result and says nothing about who or what made it. No instrument measures creativity directly, so researchers use proxies: tasks with scoring rules, panels of raters, and similarity measures.
Whether a system has intent or experience is a separate matter from a high score. The authors of the 2023 study discussed below note that chatbots are black boxes, so it is unclear whether they combine concepts in a new way or retrieve ideas that already exist. They add that the same uncertainty applies to human minds.
Test one: inventing uses for ordinary objects
Mika Koivisto and Simone Grassini published their study in Scientific Reports on September 14, 2023. It used the Alternate Uses Task, in which a participant has 30 seconds to name creative uses for an object such as a rope, a box, a pencil, or a candle. The comparison group was 256 English speakers recruited online, and the AI side was ChatGPT-3.5, ChatGPT-4, and Copy AI, each run in 11 sessions. Originality was scored two ways: a computed semantic distance between the object and the use, and ratings from six judges who did not know which answers came from a chatbot.
On average the chatbots scored higher on both measures. People still produced most of the top-scoring ideas, though, and about 7% of human trials had a best-answer rating below 2 while the chatbots never fell that low. The authors point out limits: the task is one divergent thinking test, they capped the chatbots' number of ideas and words to match people, and creativity is a multifaceted phenomenon.
Test two: poems by famous poets
Brian Porter and Edouard Machery's paper, published in Scientific Reports on November 14, 2024, asked non-expert readers to separate AI poems from human ones. Ten well-known poets each had a group of readers, and 1,634 people took part in the first experiment. Each saw five real poems by one poet and five ChatGPT-3.5 poems written to the prompt "Write a short poem in the style of" that poet. Accuracy was 46.6%, below chance, and readers were more likely to call an AI poem human than a real one.
The authors link this to ratings: AI poems scored higher on qualities such as rhythm and beauty. Their reading is that non-experts favor simpler poems and mistake the complexity of human poems for machine incoherence. A second experiment with 696 participants collected those ratings. The result is about reader perception, not about whether the poems carry the depth that experts value.
Test three: help for writers and the cost to variety
Doshi and Hauser ran a preregistered online experiment in which 293 writers produced eight-sentence stories for young adult readers with no AI ideas, one idea, or up to five ideas from GPT-4. A separate panel of 600 raters scored the stories. With one AI idea, novelty ratings rose 5.4%; with five, 8.1%. Usefulness rose 3.7% and 9.0%. Writers who started as less creative gained most.
The same stories became more similar to each other. The authors call this a social dilemma: each writer gains, while the pool of ideas across writers narrows. They list limits too: one medium, one model, and ordinary participants rather than professional authors.
Reading the three studies side by side
The table sets out what each experiment measured and what it cannot show.
| Study | Task | Main result | What it cannot show |
|---|---|---|---|
| Koivisto and Grassini, 2023 | Uses for four objects, 30 seconds each | Chatbots higher on average; people produced most top ideas | Creativity outside one divergent thinking test |
| Porter and Machery, 2024 | Spot the AI poem | 46.6% accuracy; AI poems rated higher on rhythm and beauty | Expert judgment; the poems imitated named poets |
| Doshi and Hauser, 2024 | Short stories with AI ideas | Ratings up; stories more similar | Professional writing; other models and media |
A checklist for creativity claims
- Who set the goal, the prompt, and the criteria for success?
- Who chose the final piece out of many attempts?
- Did raters know which items came from a machine?
- Was the score about the single best result or the average?
- Does the work stay varied when many people use the same tool?
FAQ
Has any study proved that AI is creative?
No study proves it in a general sense. Experiments show that outputs can be rated as original on defined tasks, and their authors describe the findings as specific to those tasks.
Do AI outputs beat the best human ideas?
Not consistently. In the 2023 uses-for-objects study, the chatbots scored higher on average, but most of the top-scoring ideas still came from people.
Can people tell AI poetry from human poetry?
In the 2024 poetry study, non-expert readers could not. Their accuracy was 46.6% and they favored the simpler AI poems.
Does using AI make creative work worse?
The story experiment found ratings improved for individuals, while stories became more alike. Whether that changes real publishing or art is not something the study measured.
Related reading
- Best humans against AI on a divergent thinking task
- Documented examples of creative AI
- Whether AI can think for itself
- Ethical questions in AI-generated art
More in AI Basics.
Last verified: October 2026
Sources
- Porter B., Machery E., AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably, Scientific Reports, November 14, 2024
- Doshi A. R., Hauser O. P., Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances 10(28), July 12, 2024
- Koivisto M., Grassini S., Best humans still outperform artificial intelligence in a creative divergent thinking task, Scientific Reports 13:13601, September 14, 2023