Using AI gives researchers a career boost – while quietly reshaping science itself

AI is rapidly transforming how research is carried out. But beyond changing methods, it may also be reshaping the careers of scientists – and may even influence what kinds of scientific questions get asked.

A new study published in Nature used computational methods to compare the research output and career trajectories of scientists who adopted AI-based techniques with those of comparable researchers who did not.

The researchers analysed 42 million scientific articles published over the past four decades to track the emergence of AI-related methods across disciplines – identifying articles that used machine learning and related techniques and linking them to publication output, citation impact and career progression. They found that AI significantly boosted the careers of some scientists in the natural sciences while subtly steering attention and resources toward certain kinds of scientific problems.

Across the dataset, researchers who adopted AI methods published roughly three times more articles, received nearly five times more citations and progressed from junior to senior researcher about a year faster than comparable peers – differences large enough to create powerful incentives that may influence which methods and problems scientists choose to pursue.

This means that using AI has been a clear strategic advantage, says co-author James Evans, a sociologist and data and AI scientist at the University of Chicago in the United States. “Individual scientists have to deal with survival – getting the next grant and getting promoted,” he says.

AI has led to tremendous strides in specific disciplines – such as AlphaFold’s Nobel-winning advances in predicting protein structure – but may also be narrowing the overall focus of science by drawing attention toward problems that are easiest to solve with large datasets and computation, Evans says.

Why AI pulls scientists toward data-rich fields

AI excels at pulling “needles from haystacks,” Evans explains, identifying subtle statistical patterns in large datasets that would be difficult for humans to detect unaided. As a result, AI adopters tend to concentrate in disciplines that already produce large, structured datasets suitable for machine-learning analysis, drawing even more resources into areas that are already well studied and potentially crowding out emerging fields that lack such datasets.

“It is fine if some people do it, but if everybody does it, then there will be diminishing marginal returns – everyone’s going to compete over smaller and smaller performance advantages until the differences are meaningless,” he adds.

Evans emphasises that this is AI’s recent history, not its destiny. “This is where AI is going, but it does not have to go there,” Evans says. “Let’s course correct.” Thoughtful funding decisions and policy tweaks at the university and governmental levels could steer resources toward both data collection – to lay the foundation for future AI-assisted research – and areas of science that cannot yet be easily optimised with AI.

“I think it would be tragic if the scientists who did not use AI were excluded – they are working on a wider set of projects and problems,” he says.

Frontier science begins with questions

So far, AI-assisted research produces good answers. But Evans says what is needed on the frontiers of science – the quest for the origins of life, the beginning of the universe or as-yet-undiscovered systems in our own bodies – are good questions rather than ever-better pattern recognition.

“AI tools are closing off or ‘completing’ fields much faster than they are generating new ones, new questions or new arguments,” he says.

“AI is not yet prepared to grapple with those classes of problems,” Evans says. But that may not remain the case for long.

In their analysis, the researchers identified three broad phases of AI adoption in science: an early machine-learning phase beginning in the 1980s, when statistical models assisted specific analytical tasks; a deep-learning phase beginning in around 2015, when multilayered neural networks enabled more complex pattern recognition; and a recent generative-AI phase beginning in around 2022.

The authors argue that a further transition could occur if artificial general intelligence emerges. In that scenario, AI systems might move beyond narrow analytical tasks toward acting as collaborators in parts of the scientific process.

From analytical tool to potential research collaborator

This generalised AI does not have to be a genius to make a difference. “Sometimes in my lectures, I ask the group to raise a hand if they think AlphaFold would help with their research projects,” says co-author Fengli Xu, Professor of Information Science at Tsinghua University in Beijing, China. Despite AlphaFold’s “superhuman” achievements, not many hands go up – predicting that protein structures remains a niche task.

“But if I change my question and say, ‘If I give you a junior PhD student, do you think it will be helpful?’ Maybe 80 to 90% say yes.”

Evans, who also leads a Google team exploring collective intelligence, says he is intrigued by developing AI research collaborators.

“How do I spin up the right AI interlocutor, conversational partner, creative partner?”

Testing whether AI can act like a scientist

In November 2025, Evans and Xu hosted a conference for leading thinkers on AI scientists in Beijing and ran a “social experiment,” Xu says. “Can AI join the conference like human scientists, with a degree of autonomy?”

To put the latest AI through its paces, they held a panel discussion between two “digital scientists” and two humans. It was not exactly a level playing field, since the humans happened to be Nobel laureates – Arieh Warshel, professor at the University of Southern California and winner of the Nobel Prize in Chemistry in 2013, and Thomas Sargent, professor at New York University and winner of the 2011 Nobel Prize in Economics.

The two AI agents, AI Specialist Lambda and AI Generalist Omega, were built using a framework called OmniScientist, which aims to endow AI with some of the soft skills needed for collaboration with other researchers. Lambda appeared on a life-size screen as a handsome young Asian man in glasses and a sweater vest, digital hands resting on crossed digital knees. Omega presented as a stern-looking grey-haired man in a bow tie. On this occasion, the AIs struggled to hold their own in conversation. “It did not go as well as I would want it to,” Xu says. But Evans hopes that after ironing out some sound issues, next year’s showing should be more impressive.

“We will have to build curiosity, different perspectives and disagreement into them, or they will suffer the same collective benchmarking fate that humans using AI tools are experiencing right now,” Evans says – reproducing the same narrowing dynamics that are already beginning to reshape science itself.