Sometimes I receive comments that I am too critical of AI. But let me make one thing clear: I try to look at both the advantages and the disadvantages. And that is not limited to AI. That is also true for this new study in Nature by Qianyue Hao and colleagues. They investigate, on a large scale, how AI influences scientific research. Their conclusion sounds both impressive and uncomfortable: AI seems to make individual scientists more successful (a benefit), while at the same time making science collectively somewhat narrower (a disadvantage).

One important clarification from the start: this research is not about papers written by ChatGPT, but about actual scientific research. The researchers primarily examine scientists who use AI as a tool in their work. Think of machine learning for medical imaging, AI for protein structures such as AlphaFold, predictive models, pattern recognition in large datasets, or generative AI to support research. In other words, this is about AI-assisted science rather than AI-written science.

The scale of the research is impressive. The authors analysed more than 41 million scientific articles from biology, medicine, chemistry, physics, materials science, and geology between 1980 and 2025. Using a trained language model, which feels oddly appropriate here, they attempted to detect which papers explicitly used AI.

At first sight, what they found is difficult to ignore. Researchers using AI publish, on average, three times as many papers, receive nearly five times as many citations, and become research leaders faster. Moreover, AI-related papers appear more frequently in top journals and receive more attention on average.

That almost sounds like a promotional brochure for AI in science. However, for me, the most interesting part of the article actually begins after that. According to the authors, AI simultaneously contributes to a narrowing of scientific exploration. To measure this, they place papers within a kind of semantic space: articles that are very similar in content are positioned closer together.

More specifically, their analysis suggests that AI research concentrates more strongly around existing, popular, and especially data-rich lines of research. “Knowledge extent”, a measure of how broadly scientific attention spreads across topics, appears to be smaller in AI-related research. Additionally, they observe less interaction among subsequent lines of research and a greater concentration of attention on a limited number of influential papers.

Today, AI seems primarily good at optimising existing research paths, not necessarily at opening up entirely new directions. Or, as the authors themselves put it: AI currently seems to automate existing fields rather than explore new ones.

Nevertheless, caution remains important. This is still observational scientometric research. The fact that researchers using AI are more successful does not automatically mean that AI is the cause. It could just as easily be that strong labs adopt AI more quickly, have more resources available, and were already more visible to begin with. The authors attempt to statistically correct for this, but openly acknowledge that causality remains difficult to establish completely.

At the same time, their measure of the “narrowing of science” also deserves nuance. Less diversity across research themes is not necessarily a bad thing. Sometimes it may simply indicate that a field is becoming more mature, collaborating more efficiently, or temporarily focusing on a breakthrough. Science, after all, does not always progress evenly across all possible ideas.

And there is another important point here. The article primarily reflects the current type of AI use in science. Today, many AI applications focus on pattern recognition, classification, prediction, and analysis of existing datasets. The authors themselves point out that AI currently works especially well in data-rich domains. That does not necessarily have to remain the case.

Still, I think the article exposes an important tension. Science today strongly rewards what is visible, scalable, and quickly impactful. AI seems to reinforce those mechanisms even further. That may lead to more efficient science, but possibly also to greater concentration around the same popular questions, datasets, and research agendas.