Is generative AI quietly making our ideas look alike?

There is a quiet paradox in using AI to be more creative. The tool can make any one person feel sharper and faster, full of options they would not have reached alone. But zoom out to the level of a whole group, or a whole industry, and a different picture may be forming. When everyone brainstorms with the same models, the ideas start to converge.

That is the conclusion of a systematic review and meta-analysis by Alwin de Rooij, at Tilburg University and Avans University of Applied Sciences, with Michael Mose Biskjaer of Aarhus University. They pulled together 19 experiments published between 2022 and early 2026, yielding 61 separate effect sizes, all comparing what happens when people create with generative software versus on their own [1].

The finding, in plain terms

Across those studies, a statistically significant homogenization effect held up [1]. People working with generative AI tended to produce ideas that were more similar to one another than people working unassisted. The researchers measured this in a couple of ways, including the semantic distance between ideas and judgments from expert raters, so it was not resting on a single definition of "alike."

The effect was not uniform, and the texture matters. It was strongest in idea-generation tasks, the kind with a fairly constrained problem where you are reaching for options. It was weakest in open-ended divergent-thinking exercises. And in a real-world slice of the analysis, the team still saw a small but significant drop in the diversity of ideas. Notably, the sameness lingered: people who had worked with AI carried some of that convergence into later creative tasks done without it.

Why it happens

The mechanism is not mysterious, and that is part of what makes it worth taking seriously. Large language models are trained on overlapping piles of internet text, and they gravitate toward the most common associations between words and concepts. Ask thousands of people to spark ideas with the same model and the model functions, in the authors' phrase, as a semantic anchor. It gently tugs everyone toward the same typical concepts, and the collective range of thinking narrows.

So the individual experience and the group outcome can point in opposite directions at once. You feel more inventive because the model hands you more than you had. The group becomes less inventive because it is being handed roughly the same things.

How firmly to hold this

A few caveats are worth stating plainly. The analysis leaned heavily on text-based language tools and did not cover every kind of AI, such as adaptive systems or music tools, so the conclusion is about a specific and very popular slice of the technology. The real-world evidence and the questions about long-term persistence still rest on a limited number of studies. The authors themselves frame the takeaway as tentative, pending broader research, and the headline study is currently a preprint rather than a peer-reviewed final paper.

None of that erases the signal. It just sets the right expectation. The useful question this raises is not whether to use these tools, but how. If a model nudges everyone toward the middle, the creative move is to treat its output as a starting point to argue with rather than a destination to settle on. Diversity of thought, it seems, is something we may now have to protect on purpose. The homogenization risk pairs with a separate finding about what happens when models are tuned for personality: research on the accuracy cost of training language models to sound warmer shows another place where optimizing for one quality quietly degrades another. And for a different angle on how we read AI behavior, there is work on what makes people attribute a moral mind to a chatbot — it turns out we are watching for contextual fit, not just tone. For more on the evolving relationship between AI systems and human judgment, see more artificial-intelligence coverage.

Sources

1. de Rooij, A., & Biskjaer, M. M. (2026). Does generative AI make us think alike? A systematic review and meta-analysis of homogenization effects in human–AI co-creation. Preprint. https://doi.org/10.31234/osf.io/rz5s4_v1

This article summarizes published research for general informational purposes only and does not constitute professional advice.