AI is turning researchers into publishing machines while quietly steering them into the same crowded corners of science. That’s the finding from a sweeping analysis of more than 40 million academic papers, reported by Hacker News, which shows that scientists who use AI tools publish more, collect more citations, and reach leadership roles faster than peers who don’t. The catch is what happens to science as a whole: it gets narrower.
The study, led by University of Chicago sociologist James Evans with collaborators at the Beijing National Research Center for Information Science and Technology, was published 14 January in Nature. What stands out here is the tension it exposes between what’s good for a scientist’s career and what’s good for scientific progress.
📊 The numbers
The team trained a natural language model to spot AI-augmented research across six natural science fields: biology, chemistry, physics, medicine, materials science, and geology. Their dataset covered 41.3 million English-language papers from 1980 to 2025. They deliberately left out computer science and math, since those fields build AI methods rather than just use them. Then they compared roughly 311,000 AI-assisted papers against millions that used no AI.
For individual researchers, AI is a rocket. On average, scientists who adopt it:
- Publish 3x as many papers
- Earn nearly 5x as many citations
- Become team leaders one to two years sooner
The collective picture is the opposite. When the researchers mapped papers in a high-dimensional “knowledge space,” AI-heavy work occupied a smaller intellectual footprint, clustered tightly around popular data-rich problems, and generated weaker networks of follow-on engagement between studies. The pattern held across four decades, from early machine learning through deep learning to today’s generative wave. “If anything,” Evans notes, “it’s intensifying.”
🔍 Why it’s happening
Evans has spent more than a decade tracking how ideas spread and stall. Back in 2008, he showed that online publishing and search pushed scientists to cite the same highly visible papers, speeding up how ideas travel but shrinking the range of ideas in play. The new work suggests AI puts that dynamic into overdrive.
The mechanism is straightforward. Models trained on abundant existing data are great at optimizing well-defined problems: predicting protein structures, classifying images, pulling patterns from massive datasets. They’re far less likely to wander into poorly mapped territory where data is thin and questions are messy. So AI ends up automating the most tractable parts of science rather than expanding its frontier. As physicist Luís Nunes Amaral of Northwestern puts it, “We are digging the same hole deeper and deeper.”
There’s a second problem. Cheap manuscript generation has fueled a surge in low-quality and fraudulent papers, produced at industrial scale. Amaral, who documented AI-fueled paper mills last year, warns that the field’s obsession with paper counts is crowding out harder questions about what the research actually contributes.
💡 What this means for you
The strategic takeaway for anyone doing or funding research:
- Use AI for what it’s good at. Optimization, pattern extraction, and well-scoped prediction are where it delivers real leverage.
- Guard your originality budget. If AI handles the tractable work, protect time for the risky, poorly mapped questions it can’t touch. That’s where differentiation lives.
- Judge work by contribution, not volume. A 3x jump in output doesn’t mean 3x the insight. Reviewers, editors, and hiring committees who reward novelty over raw counts push against the conformity loop.
- Watch the co-scientist trend. Some systems now propose hypotheses and new directions. But Evans cautions they’ll only explore uncharted ground if they’re deliberately designed and incentivized to do so.
A fair limitation to name: the study measures correlation across a huge corpus, not a controlled experiment, and it focuses on natural sciences. It can’t prove AI causes narrowing on its own, since career incentives were already nudging scientists toward safer, crowded questions long before ChatGPT.
Still, the signal is hard to ignore. The danger Evans describes isn’t that science slows down. It’s that it becomes more homogeneous while individual labs race ahead. As AI co-scientists get more capable, the open question is whether anyone will design them to reward surprise. Full details are available at the original source.