Better Ideas, Fewer of Them

Ask a generative AI model to help with a creative task and, by most measures researchers use, the result is better. Ideas are rated more novel, stories score higher with outside readers, brainstorms produce more usable output per person. This finding has held up across dozens of studies and is not seriously disputed. It is also, on its own, a misleadingly happy summary, because it describes what happens to one person's output and says nothing about what happens when millions of people run the same tool over the same kind of task at the same time.

A 2026 meta-analysis addresses exactly that gap, and it is the most rigorous accounting yet of a pattern this site has previously described only in passing. Researchers Alwin de Rooij and Michael Mose Biskjaer pooled 19 empirical studies — 61 separate effect sizes, spanning divergent thinking, idea generation, writing, and visual art, all published since generative AI tools went public — and found a statistically significant homogenization effect running through the creative-task literature as a whole. It was strongest in idea generation specifically, the task most directly upstream of everything else a creative process produces. Real-world data, not just lab experiments, showed a small but significant reduction in idea diversity, and the effect appeared to persist beyond the moments people were actively using the tool — meaning the influence doesn't switch off when the chat window closes.

The same shape, higher stakes

That would be a notable finding on its own. What makes it worth a longer look is a separate paper, published in Communications Psychology by researchers at Copenhagen Business School and the University of Cambridge, tracing an almost identical mechanism into a domain where narrowing carries very different consequences: scientific research itself. Cecilie Steenbuch Traberg, Jon Roozenbeek, and Sander van der Linden describe a self-reinforcing loop — AI-augmented papers get cited more, AI-adopting researchers publish more often, and both effects feed back into which methods, topics, and even turns of phrase get treated as the legitimate, fundable way to do the work. Evidence cited in the piece shows AI adoption already tracking with reduced topical breadth in published research. The authors' specific worry is not that AI tools produce bad science — mostly they don't — but that "topics that do not fit easily within the AI narrative may struggle for legitimacy or resources," which is a much quieter kind of narrowing than outright bias, and correspondingly harder to notice from inside it.

It's worth pausing on why this is a different problem than a slightly-less-original short story. A culture's fiction can afford some convergence; a culture's science depends on maintaining approaches that current tools and current funding logic don't happen to reward yet, because the next genuinely useful idea is disproportionately likely to come from exactly the place nobody was already looking. A field that quietly stops funding the methodologically unfashionable isn't making an error in any single decision. It's shrinking the space it draws its future ideas from, one ordinary-looking grant cycle at a time.

Why this isn't a fixed property of the technology

The most useful finding in this year's research, though, may be the one that complicates the doom narrative rather than extending it. A separate study on human-AI collaborative storytelling set out to test whether the creativity-diversity trade-off — good for the individual, bad for the group — was inherent to how these models work, or just an artifact of how they're typically deployed. The researchers built ten distinct AI "personas," each combining randomly varied traits across fourteen dimensions — cultural background, cognitive style, professional interest, literary taste — and had each one generate story plots independently. When 100 participants wrote short stories with plots drawn from this deliberately diversified pool, the homogenization effect didn't just shrink. It reversed: AI-assisted stories showed significantly less similarity to the group average than stories written with no AI help at all, while still gaining the richer vocabulary and description that AI assistance usually provides.

The mechanism the researchers identify is almost banal once stated plainly: when many people are fed highly similar creative stimuli, their outputs gravitate toward similar territory, because that's what shared stimuli do to any group of minds, AI-generated or not. The homogenization documented across the broader literature isn't a mysterious property of machine intelligence. It's closer to an old, familiar phenomenon — a monoculture forming around a single dominant input — that happens to be moving unusually fast because the input in question now reaches hundreds of millions of people through a small number of underlying models.

What doesn't follow, and what does

None of this means the fix is simple, and the science-monoculture paper is explicit that the remedy runs against the current incentive structure rather than with it: editorial boards would need to deliberately protect methodological pluralism, and funding bodies would need to resist rewarding whatever approach is currently easiest to publish fast. But the persona study does establish something the more alarming framing of this research tends to skip past — that convergence is a consequence of a specific, common deployment choice, not a ceiling built into the technology. A tool that hands a billion people the same default voice will average those people toward each other. A tool designed, deliberately, to hand different people different starting points does not have to.

The uncomfortable part is that the current default is the first kind, everywhere, by far — not because anyone chose sameness on purpose, but because sameness is what a single dominant model, offered the same way to everyone, produces by default. Fixing that isn't a matter of using AI less. It's a matter of noticing that the setting was never neutral to begin with.

Sources

1. Real-world evidence shows generative AI is making human creative output more uniform — PsyPost, reporting on de Rooij & Biskjaer meta-analysis (Journalism)
2. Generative AI Makes Creative Output More Homogeneous — Proceedings of the 2026 European Conference on Cognitive Ergonomics (Peer-Reviewed)
3. AI is turning research into a scientific monoculture — Communications Psychology, Copenhagen Business School / University of Cambridge (Peer-Reviewed)
4. Diverse AI personas can mitigate the homogenization effect in human-AI collaborative ideation — (Peer-Reviewed)
5. Metacognition can mitigate AI-driven homogenization of ideas — Nature Reviews Psychology (Peer-Reviewed)