In January 2026, Nature published a study that put numbers on something many scientists have felt. Qianyue Hao, James Evans and colleagues used a language model to flag papers that relied on AI tools, across 41.3 million research papers in the natural sciences [1]. Scientists who did AI-augmented research published 3.02 times more papers, received 4.84 times more citations, and became research project leaders 1.37 years earlier than those who did not [1].
That is the good news for the individual. The other half of the finding points the opposite way. AI adoption was associated with a 4.63% shrinking of the collective volume of topics studied, and a 22% drop in how much scientists engaged with one another’s work [1].
My reading is that this is a real tension, and an uncomfortable one for people like me. The study does not prove that AI causes either outcome. It does suggest that what is rational for each scientist and what is healthy for science can pull apart, and we should stop assuming the two always line up.
What the study measured
The team looked at six fields: biology, medicine, chemistry, physics, materials science and geology, with papers from 1980 to 2025 [1]. They deliberately left out computer science and mathematics, where AI methods themselves are built, so they could focus on science that uses AI as a tool [1].
To measure breadth, they placed each paper on a kind of mathematical map built from its text, then measured how much ground a batch of papers covered [1]. AI-augmented papers covered less of that map, and the pattern held in all six disciplines [1].
The second measure is about conversation. When many papers cite the same original study, do they also cite each other? For AI papers, less often: about 22% less [1]. The authors describe a star shape around specific popular research topics, instead of a web of connected work, and call these clusters lonely crowds [1].
Why data may be steering the work
The authors write that AI-augmented work moves collectively “towards areas richest in data” [1]. In their extra analyses, funding priority and a topic’s original impact looked almost unrelated to where AI was adopted, while data availability stood out as a major factor [1].
That makes sense to anyone who has trained a model. A deep-learning tool needs lots of examples. If a question already has a large public dataset, AI can be pointed at it next week. If the question needs data nobody has collected yet, it waits.
Imagine two graduate students starting on the same day. One picks a problem with a big public image dataset, trains a model, and has a paper out within a year. The other picks a question with no dataset and spends that year building equipment and collecting samples. On paper, the first student looks faster, even if the second one is asking the newer question.
The authors offer a useful picture of what happens when thousands of scientists make the first student’s choice. They call it collective hill-climbing: many climbers crowd the same popular mountain by the same route, which may discourage anyone from looking for higher peaks elsewhere [1]. Their summary is that AI tools “seem to automate established fields rather than explore new ones” [1].
A pattern, not a verdict
How we read this matters. The study is observational: it compares papers and careers that used AI with those that did not, after the fact. The authors write that “we cannot fully identify the causal linkage between AI adoption and scientific impact” [1]. Their method can also miss AI use that a paper never mentions [1].
There are obvious alternative stories. Early adopters may be more ambitious, better funded, or already in busy fields. The authors did compare scientists with similar early-career positions and report that the advantages still held, which they read as a sign that AI itself contributes [1]. I find that suggestive, not settling. Matching people on where they started cannot capture everything.
The narrowing result carries a similar caveat. A smaller spread of topics in AI papers could mean AI pulls people toward crowded areas. It could also mean crowded, data-rich areas are simply where AI is easiest to use. Either way, the effect on the overall map of science looks much the same, which is why I think the finding matters even without clean proof of cause.
The case for concentration
There is a fair counterargument, and the authors make part of it themselves. More overlapping attention “may benefit scientific replication and extension,” helping solid, practical answers to specific questions emerge faster [1]. A narrower focus is not automatically waste.
I agree with that, up to a point. Hard practical problems deserve concentrated effort, and medicine has plenty of them. My worry is balance. The same paper warns that when attention piles onto the same developments, science becomes more likely to get stuck on a local peak instead of searching more widely [1].
An old problem that AI may speed up
This tension did not start with AI. In a 2021 analysis, Johan Chu and James Evans found that when a field publishes a very large number of papers each year, citations flow disproportionately to papers that are already well cited, and new papers rarely break through [2]. They concluded that the progress of large fields “may be slowed, trapped in existing canon” [2]. They also noted that paper counts help decide scholars’ careers [2].
In 2024, Lisa Messeri and M. J. Crockett warned in Nature that AI tools can give scientists an illusion of understanding, which makes it harder to notice scientific monocultures forming, where some methods, questions and viewpoints come to dominate the alternatives [3]. They argued that AI risks a phase of science “in which we produce more but understand less” [3]. The 2026 study does not test that argument directly, but its numbers fit the worry better than I would like.
What I would change
This one is personal. I am a researcher who builds deep-learning tools for breast MRI and large language model tools that pull structured information out of medical records, so I feel the pull of fast, data-rich AI papers on my own career.
I do not think the answer is for individual scientists to avoid AI. The fix belongs mostly with the people who set the incentives. Funders and hiring committees could give real credit for the slow work of collecting new kinds of data, and not only for new analyses of data that already exists. The authors make a related point: they call for AI systems that help scientists gather new types of data from places we could not reach before, rather than mainly squeezing more out of the data we already have [1].
AI can be good for your career and still narrow what science asks, and this study suggests both may be happening at once. I would rather we change what we reward than ask each young scientist to choose between their CV and the harder question.
For readers, the practical habit is to notice what kind of question an AI study is answering. Another model on a famous dataset can be useful, but it is different from opening a new area. For scientists, especially those of us early in our careers, a good rule is to keep at least one project that starts with a question rather than a dataset.
References
[1] Q. Hao, F. Xu, Y. Li, and J. Evans, “Artificial intelligence tools expand scientists’ impact but contract science’s focus,” Nature, vol. 649, no. 8099, pp. 1237-1243, Jan. 2026, doi: 10.1038/s41586-025-09922-y.
[2] J. S. G. Chu and J. A. Evans, “Slowed canonical progress in large fields of science,” Proceedings of the National Academy of Sciences, vol. 118, no. 41, Art. no. e2021636118, Oct. 2021, doi: 10.1073/pnas.2021636118.
[3] L. Messeri and M. J. Crockett, “Artificial intelligence and illusions of understanding in scientific research,” Nature, vol. 627, no. 8002, pp. 49-58, Mar. 2024, doi: 10.1038/s41586-024-07146-0.
Related Articles
