Not for the person using it, on the experiments that exist, and possibly for everyone else. When writers and idea-generators are given AI help, the work is rated more creative, most of all for those who start out less creative, and different people's work comes out more alike. Each writer gains and the shared pool of ideas gets narrower. All of this comes from short tasks, mainly story writing and brainstorming, and nobody has measured what happens to a working creative's range across years.
The answer, in one line
On the experiments so far, it helps the individual and narrows the group. Writers given AI ideas were rated more creative, most of all the less creative ones, and their stories were more alike.
Definition#
Creative homogenisation: the narrowing of the range of ideas or work produced across many people when they draw on the same AI system, even while each person's own output improves. The term is descriptive, used in this sense by Anderson, Shah and Kreminski, and by Doshi and Hauser as a fall in collective diversity. It is not a SuperSkills coinage.
Each writer gains, and the shelf of stories narrows#
Doshi and Hauser, in Science Advances in July 2024, had roughly 300 writers produce short stories, some with ideas supplied by a language model, and had 600 evaluators rate them. Stories written with AI ideas were rated more creative, better written and more enjoyable, and the gains were largest for the writers who were less creative to begin with. The stories were also more similar to one another. The authors call the result a social dilemma: every writer is better off using the tool, and the combined output of all of them covers less ground.
A second experiment found the same convergence and a weaker sense of ownership#
Anderson, Shah and Kreminski compared ChatGPT with an alternative creativity support tool in a study of 36 participants, presented at the Creativity and Cognition conference in 2024. Different users produced less semantically distinct ideas with ChatGPT than with the other tool. ChatGPT users generated more ideas and more detailed ones, and felt less responsible for them. Thirty-six people is a small sample, so the direction is better supported than any size of effect.
The narrowing sits across models as well as inside one#
A reasonable objection to both experiments is that each used a single model, and a different model might spread people out. Wenger and Kenett tested that in PNAS Nexus in March 2026. They gave three standard creativity tasks to 102 people and 22 models and compared how much responses varied within each group. Models resembled one another far more than people resembled other people: on the Alternative Uses Task the variability score was 0.459 for models and 0.699 for people, with the same ordering on the other two tasks. This compares models with people. It does not follow anyone who worked with a model, and the authors themselves say only that using models as creative partners may push users towards similar outputs.
Three limits on how far these experiments reach#
- Short tasks. Eight-sentence stories and brainstorming sessions are where a tool that supplies a competent first idea has the most to offer. A film, a campaign or a product has many more decisions in it.
- Strangers as judges. Evaluators rated work on the page. Nothing here measures whether a market, a client or an audience rewards the more individual version.
- One session. No study follows people for months to see whether their unaided range shrinks, recovers or does not move. The weaker sense of ownership in the Anderson study is the nearest thing to a signal, and it measures a feeling and not a skill.
A rising floor stops telling people apart#
This section is interpretation, kept apart from the evidence above.
The largest gains in the Doshi and Hauser experiment went to the people who started lowest. A tool that lifts the weakest work most and gives everyone similar material to start from is a floor, and when everyone stands on the same floor it stops distinguishing anyone. Rahim Hirji describes the practical version in Solving Synthetic Seniority (5 July 2026), where a studio head takes laptops away from her graduates and sets the same brief to a designer who works with the tool. The designer finishes first and the graduates' hand-made work is better. His line for the tool is that it cannot fail you and cannot lift you. The seventy per cent he puts on where that floor sits is his own estimate, offered in an essay, and no study measures it. The studio is one account, from one firm, in one trade.
What the experiments would predict from such a floor is the pattern the essay describes: acceptable work arrives quickly, the effort that used to build the last stretch of quality is skipped, and the people who needed that effort most are the ones the tool helps most. That is a reading of the evidence, and a long-run study of working creatives would test it. The related question of whether readers can tell is taken up in what is the human signal.
Keeping the tool away from the work that carries your name#
- Sort the work by whose name is on it. Meeting notes and a first-pass email can go to the tool at no cost. The pitch, the design or the article that a client buys because it is yours is where a shared floor costs the most.
- Make a version before you ask. In the experiments, access to AI ideas was what moved people towards one another. Writing your own rough version first, then asking the tool to challenge it, keeps your starting point yours. This is a hypothesis drawn from those results and not a tested remedy.
- Compare outputs across a team. If five colleagues give the same tool the same brief and the answers could be swapped, the range has already narrowed. The shared prompt review sets out one way to check.
- Protect a manual rep. Where the tool lifts the weakest skills most, those are the skills that most need practice. See how to keep your own voice when using AI.
Key sources
- Doshi, A. R. and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances 10(28), eadn5290. Graded entry.
- Anderson, B. R., Shah, J. H. and Kreminski, M. (2024). Homogenization Effects of Large Language Models on Human Creative Ideation. Creativity and Cognition 2024, DOI 10.1145/3635636.3656204. Graded entry.
- Wenger, E. and Kenett, Y. N. (2026). Large language models are homogeneously creative. PNAS Nexus 5(3), pgag042. Graded entry.
Related SuperSkills research#
On the population-level evidence, does AI make everyone think alike. On what readers look for, what is the human signal. On output that outruns the judgement behind it, synthetic seniority. On what machines leave to people, what stays human.
About this research#
Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The three studies were read at source on 1 October 2026 and are graded in the evidence base. The Science Advances paper itself returned a 403, so its sample sizes are the authors' institutional figures and its findings are taken from the repository abstract. Creative homogenisation is a descriptive term and not a SuperSkills coinage.
Evidence review · SS-2026-377 · Graded against the published rubric · 3 peer-reviewed studies
Hirji, R. (2026). Is AI bad for creativity?. The SuperSkills evidence base, SS-2026-377. https://thesuperskills.com/research/is-ai-bad-for-creativity. Last reviewed 1 October 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work