Artificial intelligence promises to accelerate scientific discovery and open new frontiers of inquiry. But new research from James Evans (Faculty Co-Director of Novel Intelligence; Max Palevsky Professor of Sociology & Data Science; and Director of the Knowledge Lab) and colleagues reveals how AI tools are expanding individual scientists' capabilities but narrowing the collective scope of science.
Published in Nature, the study analyzed 41.3 million research papers to find that scientists who use AI publish 3.02 times as many papers, receive 4.85 times as many citations, and even become research leaders 1.4 years earlier than those who don't. Yet collectively, AI adoption shrinks the volume of scientific topics studied overall by 4.63% and decreases engagement between scientists by 22%.
Why? Scientists using AI migrate toward areas with abundant data, where AI tools demonstrate measurable advances on scientifically legible benchmarks. Rather than expanding exploration across science, AI concentrates attention on data-rich domains while leaving a growing number of potentially fruitful areas unexplored.
The study found that AI research creates what the authors call "lonely crowds," or popular topics that attract concentrated attention but with reduced interaction among papers citing the same work. This leads to more overlapping research and a contraction in knowledge extent, as scientists converge on the same solutions to known problems rather than generating new ones.
As Evans discussed in his recent Science article "After Science," AI's efficiency risks creating methodological monocultures. Without diverse approaches, science risks premature convergence on established paradigms rather than exploring genuinely novel directions.
The research points to a need for policy intervention to actively promote new data gathering and alternative uses of AI that expand rather than contract science—including incentivizing research in data-poor areas and encouraging AI systems designed for exploration rather than optimization. The authors note how the very models that can generate highly probable outputs are also uniquely situated to recognize the character of surprising data, artifacts, and their scientific consequences.
"To preserve collective exploration in an era of AI use," the authors conclude, "we will need to reimagine AI systems that expand not only cognitive capacity but also sensory and experimental capacity, enabling and incentivizing scientists to search, select and gather new types of data from previously inaccessible domains rather than merely optimizing analysis of standing data." They argue that this capacity–to create and not merely compress data–will be required for AI to support sustainable scientific advance.
Read the full paper in Nature .
