AI makes scientists faster and more cited - but it narrows what we study
AI is now a daily collaborator in labs: planning experiments, summarizing papers, doing math, and crunching data. A large-scale analysis of 41.3 million papers across physics, biology, chemistry, geology, materials science, and medicine shows a clear trade-off: AI boosts output and impact, while shrinking the range of topics by about 5%.
If you're leading a group or building a career, this matters. You can get more done, get cited more, and move up faster - yet drift into the same crowded questions as everyone else.
What the data shows
The team used an AI language model to detect AI-assisted work within the OpenAlex database and flagged nearly 310,000 AI-augmented papers. Across disciplines, AI-supported publications drew more citations, were more impactful by multiple indicators, and appeared more often in high-impact journals.
At the individual level, researchers using AI published about three times as many papers and received almost five times as many citations as non-users. In physics, AI adopters averaged 183 citations per year versus 51 for those who didn't use AI.
Careers moved faster too. Among more than two million scientists, junior researchers who adopted AI were more likely to become established and stepped into project leadership roughly 1.5 years earlier on average.
The trade-off: convergence on well-trodden problems
When the authors examined a random sample of 10,000 papers (half AI-assisted), AI-augmented work covered a narrower spread of topics by nearly 5% and clustered more tightly. In short: AI is steering attention to problems with abundant, convenient data and mature methods.
That can push effort away from foundational questions and into operational, data-rich niches. The fix isn't to use less AI - it's to use it differently: expand what you can sense, measure, and explore, not just how fast you analyze what's already on your desk.
Use AI without narrowing your science
- Set exploration quotas: Allocate a fixed share of projects to high-uncertainty topics and track that share quarterly. Protect that time like a grant deadline.
- Prompt for divergence, not consensus: Ask models for "least-studied variables," "out-of-distribution hypotheses," and "experiments that would falsify the dominant explanation." Require at least one proposal that needs new data, not just re-analysis.
- Expand your sensing capacity: Use AI to design sampling plans, evaluate new instrumentation, or automate data collection pipelines so you can gather measurements from previously inaccessible environments or scales.
- Cross-pollinate fields: Feed the model methods from an adjacent discipline and ask, "Which of these could transfer and why?" Then validate with a domain expert before committing resources.
- Build your own datasets: Prioritize data that doesn't exist publicly. Even small, well-documented datasets can open lines of inquiry others can't follow easily.
- Track novelty explicitly: Measure topic entropy or Jaccard distance of your keywords, references, and venues. If entropy drops for two quarters, inject a new question or collaborator.
- Structure literature reviews for surprise: Force the model to surface papers with low citation overlap to your current corpus and explain why they still might be relevant.
Practical benchmarks to keep your work broad
- Topic spread: Entropy of keywords/abstracts across your last 10-20 papers or proposals.
- Citation clustering: Share of references drawn from your top five journals and top five authors; aim to reduce this over time.
- New data ratio: Percentage of results derived from data you collected this year vs. public/legacy datasets.
- Transfer attempts: Number of methods imported from other fields and tested in pilot experiments per quarter.
Signals from the field
Use is surging. One provider reports messages on advanced science and math topics approaching 8.4 million per week, up nearly 50% year over year. That lines up with what many labs feel day to day: AI is already part of the stack.
Resources
- OpenAlex - useful for mapping fields, tracking topic drift, and finding under-cited areas.
- OpenAI blog - updates on model capabilities and usage patterns relevant to research workflows.
Level up your lab's AI workflow (without losing originality)
If you're formalizing AI skills across your team - literature synthesis, experiment planning, and data analysis - structured training helps. See our role-based options here: AI courses by job.
Prefer a credential tied to analysis workflows? This path is focused on practical work with research data: AI certification for data analysis.
Bottom line
AI is a force multiplier for output, citations, and career momentum. To keep your science from collapsing into the obvious, couple that speed with intentional exploration: new data, new sensors, and explicit novelty goals.
Your membership also unlocks:
