What benefits each person does not benefit the collective.
A sweeping analysis of 41 million scientific papers, accepted by Nature, has surfaced a paradox as old as collective action itself: the tool that makes each individual more powerful is quietly making the whole enterprise smaller. Researchers who adopt AI publish more, earn more recognition, and rise faster — yet the cumulative effect is a science that explores fewer questions, crosses fewer boundaries, and concentrates its gaze on the already well-lit corners of knowledge. James Evans and his team at the University of Chicago have named what many fields are beginning to feel but have not yet measured: that optimizing for the individual can quietly impoverish the commons.
- Scientists using AI are tripling their output and advancing their careers nearly 1.4 years faster — making adoption feel less like a choice and more like a survival imperative.
- Beneath the individual gains, the scientific map is shrinking: topic diversity has fallen by 5% and cross-study citations have dropped by 22%, as research gravitates toward data-rich, already-established domains.
- AI functions not as a broad accelerant but as a magnet, pulling researchers toward where models perform well and citations are nearly guaranteed — rational for each person, corrosive in aggregate.
- The concentration of attention is stark: fewer than a quarter of all papers now capture 80% of citations, replacing the dense, interwoven network of science with a handful of orbiting stars.
- The structural risk is not confined to laboratories — journalism, law, and finance face the same mechanics, where the sum of efficient individual decisions can quietly hollow out an entire field.
A study of more than 41 million scientific papers, accepted for publication in Nature, has uncovered a paradox at the heart of AI's role in research: the tool meant to expand human knowledge may instead be narrowing it.
James Evans, a sociologist and data scientist at the University of Chicago, led a team examining papers published between 1980 and 2024 across six natural science disciplines. Of the 41.3 million papers analyzed, roughly 311,000 showed signs of AI assistance — and those papers told two contradictory stories at once.
For individual researchers, the gains are undeniable. AI-assisted scientists publish three times as many papers, receive nearly five times as many citations, and reach leadership positions over a year earlier than their peers. In a profession where funding and tenure depend on measurable output, the tool is less an option than a competitive necessity.
But what benefits each researcher costs science as a whole. AI-assisted research covers less thematic ground, accelerating work in established, data-rich domains rather than opening new frontiers. Topic diversity has shrunk by nearly five percent. More subtly, AI-produced papers generate 22 percent fewer connections between studies — they cite each other less, spawn fewer new directions, and cluster around a small number of dominant papers rather than weaving the dense, cross-referencing network that drives scientific refinement.
The mechanism is precise: AI functions as a magnet, pulling researchers toward where data already exists and citations are nearly guaranteed. The rational choice of each individual — to use the tool that maximizes their career — produces, in aggregate, a narrower science. Evans identified this as a public goods problem: what benefits each person does not benefit the collective.
The pattern is not unique to laboratories. Any system where AI optimizes individual performance — journalism, law, financial analysis — faces the same structural risk. The metrics by which we measure success remain blind to costs that never appear on any dashboard. A researcher watches their publication record grow; the questions never asked go unrecorded. What was promised as a rising tide turns out to be a current — powerful, but pushing everything toward the same destination.
A study of more than 41 million scientific papers, accepted for publication in Nature, has uncovered a paradox at the heart of artificial intelligence's role in research: the tool that was supposed to expand human knowledge is instead narrowing it.
James Evans, a sociologist and data scientist at the University of Chicago, led a team that examined papers published between 1980 and 2024 across six natural science disciplines—biology, medicine, chemistry, physics, materials science, and geology. They trained a model to identify which of the 41.3 million papers in their dataset had been assisted by AI, finding roughly 311,000. What they discovered in those papers tells two contradictory stories at once.
For individual researchers, the numbers are stark. Scientists who adopt AI publish three times as many papers as their peers who don't. They receive nearly five times as many citations. They advance to leadership positions in research teams 1.37 years earlier. In a profession where survival depends on publication counts and citation metrics, where funding and tenure hinge on demonstrating impact, a tool that triples output is not optional—it is a competitive necessity. Evans put it plainly: researchers face the same shared problem as everyone else. They need money to survive. To get money, they need to produce. The individual incentive is irresistible.
But the gains for each researcher come at a cost to science as a whole. The study reveals that AI-assisted research covers less thematic ground. Rather than opening new disciplinary frontiers, the technology accelerates work in domains that are already established and rich in data. The overall volume of topics explored shrinks by nearly five percent. There is a second, subtler effect: papers produced with AI generate 22 percent fewer connections between studies. They cite each other less frequently. They spawn fewer new research directions. Instead of building the dense, interwoven network of studies that refine one another, they orbit around a handful of star papers. Fewer than a quarter of all papers capture 80 percent of citations.
The mechanism is precise. AI does not function as an even acceleration across the scientific landscape. It functions as a magnet, pulling each researcher toward where data already exists, where models perform well, where citations are nearly guaranteed. The rational choice of each individual scientist—to use the tool that maximizes their career—produces, in aggregate, a narrower science. Evans identified this as a public goods problem: what benefits each person does not benefit the collective.
The pattern extends beyond laboratories. Any system where AI optimizes individual performance—journalism, law, financial analysis—faces the same structural risk. The sum of efficient individual decisions can impoverish an entire field. The study does not measure this spillover effect, and Evans's team did not venture beyond the six disciplines they analyzed. But the mechanics they describe contain nothing unique to science.
The research does not argue for halting AI adoption. It points to something harder to process: the metrics by which we measure success are blind to costs that never appear on any spreadsheet. A researcher watches their curriculum vitae grow. The loss of questions never gets recorded on any individual dashboard. It is a distributed system cost, with no owner, no one accountable. AI does not expand knowledge. It concentrates it. It improves the position of those who use it while simultaneously shrinking the map of what science dares to investigate. The original promise—a rising tide lifting all boats—mischaracterized the phenomenon. What exists instead is a current: powerful, but pushing everything toward the same destination.
Notable Quotes
The researcher has a shared problem: survival. To survive requires money. To get money requires production.— James Evans, University of Chicago
If all of us climb the same mountain, there are many mountains no one ascends.— Study team description of the concentration effect
The Hearth Conversation Another angle on the story
So the study shows that using AI makes individual scientists more productive. That sounds like a win. Where's the problem?
The problem is that everyone is becoming more productive in the same direction. AI doesn't help you explore new territory—it helps you mine existing territory faster. You get more papers, more citations, but the field as a whole stops asking new questions.
But couldn't a scientist just choose to work on something novel instead? Why does AI push them toward established topics?
Because the tool works best where there's already lots of data. And because the incentive system rewards volume and citations, not originality. If you use AI to work on a crowded, data-rich problem, you'll publish faster and get cited more. If you try something genuinely new, the AI can't help you as much, and you fall behind competitively.
That sounds like a problem with how we measure success in science, not with AI itself.
Exactly. AI just exposes and amplifies what was already broken. The metrics—papers, citations, speed of promotion—were always blind to diversity of thought. AI just makes that blindness more consequential. It's like giving everyone a faster car on a road that only goes one direction.
Is there a way to fix this? Can science change how it measures success?
That's the hard question. You'd have to change the entire incentive structure simultaneously across universities, funding agencies, and journals. Individual researchers can't opt out without sacrificing their careers. It's a collective action problem. Everyone knows the current path narrows the field, but no one can afford to be the first to step aside.