> A recent Nature study reveals a sharp paradox: AI tools amplify individual scientists' output while making the entire scientific enterprise narrower, more homogeneous, and more dependent on data-rich, well-established fields.
The Picture in One Image
Imagine a vast library. For centuries, scientists explored it like adventurers—some discovered untouched manuscripts on the basement level, others found dust-covered maps in the attic. Occasionally they met in the hallways, exchanged notes, and walked off in different directions.
Now AI arrives, handing each scientist night-vision goggles and an electric scooter. Everyone moves ten times faster—but almost all of them are in the same corridor.
The Data: An "AI Dividend" at the Individual Level
Tsinghua University's Hao Qianyue team, together with James Evans of the University of Chicago, analyzed 41.3 million natural science papers across six disciplines, using a fine-tuned BERT model (F1 = 0.875) to identify AI-augmented research.
Scientists using AI outperform across the board:
| Metric | AI users vs. non-users | |--------|------------------------| | Papers published per year | 3.02x | | Citations received | 4.84x | | Time to becoming PI | 1.37 years earlier | | Team size | 1.33 fewer members |
Teams shrink but output grows. Junior researchers on teams decline by 31.14% (from an average of 2.89 to 1.99), senior researchers by 10.77% (from 4.01 to 3.58). AI is automating the repetitive labor once assigned to junior staff—literature reviews, data cleaning, code debugging.
AI also reduces the risk of junior researchers leaving academia. Survival analysis (Extended Data Fig. 7) shows AI-using junior researchers are promoted to senior positions faster, while their time to exit academia is similar or slightly longer. In short: AI gives you more reasons to stay and a faster ladder to climb.
But What Happens at the Collective Level?
Collective contraction of scientific topics: 4.63%
At the scale of 41.3 million papers, this means tens of thousands of research directions are being abandoned or never explored. The knowledge scope of AI papers is markedly narrower than non-AI papers, and more than 70% of over 200 sub-fields show shrinking knowledge scope (Extended Data Figs. 8–9).
Scientist-to-scientist interaction drops 22%
Science has never been a solo game. Einstein needed the Copenhagen school's pushback; Watson and Crick needed Franklin's X-ray diffraction images. A 22% drop in interaction means fewer hallway conversations, fewer interdisciplinary collisions, and lower odds of serendipitous discovery.
Matthew effect intensifies: higher citation Gini coefficient
Extended Data Fig. 10 is stark: in AI research, roughly 20% of papers capture 80% of citations, and 50% of papers capture 95%. AI is not democratizing science—it is making it more concentrated and more winner-take-all, as AI-boosted super-papers vacuum up citations while marginal work is pushed further to the edges.
The Core Paradox
> "AI tools expand individual scientists' impact while making the scientific enterprise narrower, more homogeneous, and more dependent on already data-rich domains."
A Feynman-style analogy: imagine villagers foraging mushrooms in a forest. Before, each carried a small basket down a different path—east, west, or into an unexplored swamp. Each found few mushrooms, but the village collected many kinds. Now someone invents a "mushroom detector" that lets you pick ten times faster in known mushroom fields. Every basket gets fuller—but the whole village only harvests in one place, because the detector doesn't work in unfamiliar swamps where there's not enough data to train it.
That is AI's role in science: an accelerator of the known, not an explorer of the unknown.
Why It Matters
Many major breakthroughs came from the periphery—directions that seemed unimportant, data-poor, or even strange at the time:
- CRISPR began as obscure research on bacterial immune systems
- Quantum mechanics emerged from a minor annoyance about blackbody radiation
- Neural networks were considered a dead end by mainstream AI before 2012
Co-author James Evans published a famous paper in 2008 showing that electronic publishing was narrowing science. Eighteen years later, his team found AI repeating the trend with greater force.
The Question Left for Every AI-Empowered Researcher
When was the last time you walked down a path "without data"?
When you use AI tools, are you exploring new questions, or answering existing questions more efficiently?
When AI makes 3x the papers and 4.84x the citations possible, should we redesign evaluation metrics—not asking "how many papers did you publish," but "what new corridor did you open"?
If science collectively contracts by 4.63%, how much "efficiency loss" are we, as individuals, willing to bear for the unknown 5%?
Nature won't answer these questions for you. But your next research topic might.
References
Hao, Q., Xu, F., Li, Y. & Evans, J. (2026). Artificial intelligence tools expand scientists' impact but contract science's focus. *Nature*, 649, 1237-1243. https://doi.org/10.1038/s41586-025-09922-y
Further Reading
Evans, J. A. (2008). Electronic publication and the narrowing of science and scholarship. *Science*, 321(5887), 395-399. https://doi.org/10.1126/science.1150473