AI tools boost individual output - and may narrow what science studies
A new analysis of 41 million papers over four decades tracks a trade-off many labs can feel: researchers who use AI tools publish more and get cited more, but their work tends to cluster in data-rich niches. That's good for careers, less good for breadth of discovery.
The signal is strong. AI adopters published roughly three times more papers and earned nearly five times more citations than peers who didn't use these tools. Teams were smaller and had fewer junior members, yet early-career researchers in AI-using groups were 13% more likely to stay in academia and reached independence about 1.5 years sooner.
How the study worked
A large language model scanned 41M+ articles (1980-2025) across biology, medicine, chemistry, physics, materials science and geology to identify studies that used AI. It grouped them into three eras: traditional machine learning (1980-2014), deep learning (2015-2022), and generative AI (2023-present). Human reviewers agreed with the model's classifications in ~90% of sampled cases.
Fields like mathematics, computer science and social sciences were excluded to focus on work that used AI to do science, not to build AI itself. One caveat: if AI use wasn't described clearly in a paper, the model likely missed it, which could undercount adoption.
Career upsides, with caveats
For individuals and PIs, the incentives are obvious: more throughput, more citations, faster paths to leadership. That said, the definition of "impact" here is based on citation counts and PI status. As Molly Crockett notes, those metrics can reflect hype and funding cycles as much as research quality.
Bottom line: AI can accelerate your lab's output and visibility. Just be honest about what those metrics measure-and what they don't.
Field-level downsides
The same analysis finds adoption concentrates attention on problems with abundant data, while fundamental, low-data questions get sidelined. Some areas "at risk of extinction" simply don't lend themselves to automation or supervised learning at current scales.
Collaboration also dipped. AI-linked work was associated with 22% less engagement between scientists, creating "lonely crowds" in hot areas-many people working adjacent to each other, not together. As James Evans puts it, we're applying AI to a narrow slice of the knowledge process, biasing toward known problems over new questions.
Practical guardrails for labs and departments
- Balance the portfolio: reserve time and resources for low-data, exploratory work alongside AI-amplified projects.
- Mix methods: pair automated pipelines with theory, mechanism-first studies, and small-N experiments that stress-test assumptions.
- Protect junior development: set minimums for junior authorship and lead roles in AI-heavy projects; mentor on problem selection, not just tooling.
- Incentivize novelty: weigh venues, preregistered studies, negative results, and cross-field impact-not only citation velocity.
- Diversify data: include underrepresented systems and edge cases; audit datasets and models for brittleness and shortcut learning.
- Collaboration quotas: require external co-authors or cross-lab replications on AI-driven studies to avoid "lonely crowd" effects.
- Declare AI use: document where AI touched the pipeline (design, analysis, writing); share prompts, models, versions and guardrails.
- Fund the hard stuff: ring-fence internal funding for low-data questions and new instrumentation that expands what can be measured.
- Evaluate wisely: add field-appropriate markers (method adoption, new assays, data resources) to promotion and grant reviews.
What to watch next
Monitor whether topic diversity and cross-field citations shrink as generative tools spread. Track lab sizes, authorship patterns and junior retention in your unit. If you're formalizing skill-building on these methods, consider structured training so AI augments reasoning rather than replacing it.
Study details in Nature (DOI: 10.1038/s41586-025-09922-y)
Curated AI courses by job role
Takeaway for research leaders
AI is a throughput multiplier for individuals. Without guardrails, it can also narrow inquiry, fragment collaboration and tilt incentives toward easy-to-measure wins. Set policies now so your lab benefits from the speed without sacrificing the questions that actually move a field forward.
Your membership also unlocks:
