Is AI creative? What the experiments say

The question “is AI creative?” often draws answers of principle. The experiments published since 2023 do better: they measure. And they describe a paradox. AI raises the average creative level, especially for those who start low, but it brings outputs closer to one another. Everyone gains, while the collective loses diversity. The real question is therefore not whether a machine can have a good idea, but what becomes of originality when everyone starts from the same ideas.

The alternate uses test: AI beats the average, not the best

A classic creativity test consists in finding unexpected uses for an everyday object: a rope, a box, a pencil. It is a measure of “divergent thinking”, the ability to produce varied and original ideas.

A study published in Scientific Reports in 2023 gave this test to 256 people and to three chatbots. Result: on average, the chatbots do better than the humans. Human responses include weak ideas, whereas the chatbots generally produce more creative responses. But, according to the authors, the best human ideas still match or exceed those of the chatbots.

One point of honesty: a correction was published in February 2024, after an error in the analysis code. It affects only one passage of the results, on each participant’s best scores, which remain higher on average for AI. It changes neither the abstract nor the conclusion of the study. In other words, AI is very good at producing good ideas consistently; the peaks of creativity remain human, and rare.

Better, but the same: the diversity dilemma

A second, pre-registered experiment was published in Science Advances in July 2024. In it, 293 participants had to write an eight-sentence story. Some wrote alone; the others could ask GPT-4 for a three-sentence starting idea, or up to five ideas depending on the group. The texts were then evaluated by 600 other participants.

The authors speak of a “social dilemma”: individually, each writer is better off; collectively, the range of novel content narrows. Compared with stories written without help, those written with AI ideas show both sides of this dilemma.

Among the most creative, on the other hand, the study finds no sign that AI ideas take them beyond what they write alone.

Uniformity is the point we find most important. If all the brands in a sector, or all the sites on a topic, start from the same suggestions, they produce correct texts, and interchangeable ones.

When AI trains on AI

There is a second mechanism, a more distant one. A study published in Nature in July 2024 trained models in a loop:

According to the authors, it is the indiscriminate use of model-generated content in training that causes this loss. Yet rare cases are precisely what makes for originality.

Here too, caution is needed. The phenomenon is demonstrated in the laboratory, in a controlled training loop, and real models mix human data and generated data. In the same study, keeping 10% of the original data at each generation left only a slight degradation in performance. The correction published by the authors in March 2025 concerns only a mathematical notation. Their conclusion does not change: as the web fills up with generated text, data from genuine human interactions gains value. What this means for the web is the subject of our article on model collapse and generated content.

Who is the author?

The law, too, is looking for where to place human creativity. It is moving fast, and every statement that follows is dated.

• In the United States, the US Court of Appeals for the District of Columbia Circuit ruled in the Thaler case, on 18 March 2025, that copyright law requires a work to be created first by a human being. It adds that this requirement does not prevent a work made with the help of AI from being protected. The Supreme Court declined to hear the case on 2 March 2026, which leaves that decision in place. The Copyright Office, the US federal copyright office, set out its position in the second part of its report on AI, in January 2025: an AI output is protected only if a human author has determined enough of its expressive elements, which a simple prompt does not do, and using AI as an aid does not prevent protection.
• In Germany, the Munich I Regional Court ruled on 11 November 2025, in a case between GEMA, the German society for the collective management of music rights, and OpenAI, that memorising song lyrics in a model constitutes reproduction, and that the text and data mining exception does not apply. The court stated at the time that the judgment was not final.
• In Europe, Article 53 of the EU AI Act has required providers of general-purpose AI models, since 2 August 2025, to have a copyright compliance policy, including respect for the opt-out expressed by rights holders, and to publish a sufficiently detailed summary of the content used for training. Providers of models placed on the market before that date have until 2 August 2027 to comply. These rules concern the training of models; we cover them in our article on AI, copyright and paying authors.

The trend is clear, even if it is not complete: what is protected is the human contribution. In the United States at least, what a machine produces without human input cannot be protected.

What we do with it

The split that seems right to us: AI can produce a first draft, the human brings the angle, the lived example and the choice. It is also what an agency should be able to show you, its review process, as we explain in what an AI SEO agency should actually deliver.

This is not only a question of ethics. It is also an argument for visibility. A page that contains only what a model would have written on its own adds nothing to the answer that model already gives. Google, incidentally, invites you to ask these questions in its guidelines for creating helpful, reliable content: does your content provide genuinely original information or analysis? Does it offer more relevant information than the other pages in the search results? A generic page therefore has little reason to be cited, whether by Google or by an answer engine. Conversely, your own data, a real case, a reasoned opinion are exactly what an assistant cannot invent. That is the thread of our article on how to get cited by ChatGPT and AI engines, and more broadly of our approach to search optimisation in the age of AI.

The limits of what we know

Creativity tests measure short laboratory tasks: finding uses for an object, writing eight sentences. They say nothing about a long work, a brand campaign or art direction, where creativity depends as much on choice and coherence as on the starting idea. Models change fast, and a 2023 experiment does not necessarily describe those of 2026. Finally, the law is moving: the EU AI Act was already amended in July 2026 by Regulation (EU) 2026/1744, which does not touch the copyright obligations cited above.

What we think

Our position, signed Jim:

For someone starting from a low base, it raises the level, and that is no small thing. But the creativity that distinguishes a brand or an author lies in the gap from that average: a point of view, an experience, a deliberate choice. The more AI spreads, the rarer that gap becomes, and the more valuable it gets.

It is the same logic as for work in general, where human judgement matters more than output. We would change our mind if experiments showed that creations made with AI remain as diverse as human creations at scale, or that models regularly produce ideas that the best humans would not have had. That is not what the studies cited here say.

In practice, start from what only you know, and use AI to shape it, not to find the starting idea. It is also the principle of our SEO and GEO content engine, where the expertise comes from your team and nothing goes out without your review.

Frequently asked questions

Sources

1. A study published in Scientific Reports in 2023 (new tab)doi.org
2. A correction was published in February 2024 (new tab)doi.org
3. Published in Science Advances in July 2024 (new tab)doi.org
4. A study published in Nature in July 2024 (new tab)doi.org
5. The correction published by the authors in March 2025 (new tab)doi.org
6. The Thaler case, on 18 March 2025 (new tab)media.cadc.uscourts.gov
7. 2 March 2026 (new tab)supremecourt.gov
8. The second part of its report on AI, in January 2025 (new tab)copyright.gov
9. 11 November 2025 (new tab)justiz.bayern.de
10. Article 53 of the EU AI Act (new tab)eur-lex.europa.eu
11. Its guidelines for creating helpful, reliable content (new tab)developers.google.com
12. Regulation (EU) 2026/1744 (new tab)eur-lex.europa.eu