Scientists who use AI tools publish three times more papers and receive nearly five times more citations than those who don’t. But as more researchers adopt AI, the collective range of topics being studied is shrinking. That is the central finding of a study published in Nature on February 17, based on an analysis of 41.3 million research papers spanning four decades.
The research, led by teams at Tsinghua University and the University of Chicago, presents an uncomfortable tradeoff: AI is making individual researchers more productive while quietly narrowing the scope of science itself.
The Individual Upside
The career benefits of using AI in research are substantial and consistent across fields. Scientists who adopt AI tools publish 3.02 times more papers per year than their peers. Their work receives 4.84 times more citations. They become principal investigators or research leaders 1.37 years sooner - about 15.75% faster than the average career trajectory.
These advantages held across all six natural science disciplines studied: biology, medicine, chemistry, physics, materials science, and geology. Papers incorporating AI methods received 98.7% higher annual citations than comparable non-AI papers.
For individual researchers weighing whether to learn machine learning techniques, the data is unambiguous: AI adoption accelerates careers.
The Collective Downside
Here is where the findings get less comfortable. As AI adoption grows across disciplines, the diversity of research topics being investigated has contracted by 4.63%. Follow-on engagement between related papers - a measure of how much scientists build on each other’s work - has dropped by 22%.
The study tracked this pattern across three eras of AI: classical machine learning (1980-2014), deep learning (2015-2022), and generative AI (2023-present). The contraction appeared in more than 70% of the 200-plus scientific subfields examined, and the effect has intensified with each new wave of AI tools.
“When you have a hammer, you go around looking for nails, and that’s what AI is right now,” said Steven Salzberg, a computational biologist at Johns Hopkins University, in an NPR interview about the findings.
Why This Happens
The mechanism is straightforward. AI tools work best on problems with large, well-structured datasets. Scientists naturally gravitate toward areas where AI provides the most measurable advantages - producing what the researchers describe as “lonely crowds” of popular but increasingly similar research topics.
The result is a concentration effect. Citation distributions become more unequal (the study measured a GINI coefficient of 0.754 for AI-assisted work versus 0.690 for non-AI work). Researchers cluster around data-rich domains while questions requiring new data collection or unconventional methods receive less attention.
James Evans, a computational social scientist at the University of Chicago and one of the study’s senior authors, warned that this pattern risks creating “methodological monocultures” where entire fields converge on the same approaches and overlook alternative paths.
What This Means
The study does not argue that scientists should stop using AI. The individual benefits are real, and many AI-assisted discoveries have genuine value. But the authors point out that what is good for individual researchers is not automatically good for science.
Biology, the field with the highest surge in AI adoption (a 51.89-fold increase), also shows some of the clearest signs of topic contraction. If AI tools push researchers toward the same well-trodden problems, the field risks missing novel discoveries that come from exploring less obvious directions.
The researchers recommend three interventions: building AI systems that expand experimental and observational capacity rather than just processing existing data, redirecting attention toward foundational questions in data-poor areas, and restructuring incentives so that researchers are rewarded for exploring new territory rather than just optimizing productivity metrics.
The Fine Print
The study used a fine-tuned BERT model (F1 score: 0.875) to identify AI usage in papers, which means some uses of AI were likely missed. The analysis is limited to six natural science disciplines and does not cover social sciences, humanities, or engineering.
The researchers acknowledge they cannot fully establish causation. The correlation between AI adoption and topic narrowing is strong and consistent, but other factors - funding patterns, publication incentives, institutional pressures - may also contribute. The generative AI era findings are preliminary given the limited publication data available so far.
Still, the dataset is enormous - 41.3 million papers from the OpenAlex database covering 1980 through 2025 - and the pattern is consistent across disciplines, time periods, and analytical approaches. Whether AI is the sole cause or an amplifier of existing trends, the direction is clear: science is getting more productive per researcher and less diverse overall.