A study of more than 41 million research papers, published in Nature in January 2026, found that scientists who use AI publish three times as many papers and receive nearly five times as many citations. The same study found that AI-driven research covers a narrower range of topics and forms “lonely crowds” of weakly connected papers.
The Study
- Title: “Artificial intelligence tools expand scientists’ impact but contract science’s focus”.
- Journal: Nature, published January 14, 2026.
- Authors: Qianyue Hao, Fengli Xu and Yong Li of Tsinghua University, and James Evans of the University of Chicago, Director of the Knowledge Lab.
- Preprint: first posted to arXiv on December 10, 2024.
The Dataset
- Papers: 41,298,433 natural-science papers published between 1980 and 2025.
- Sources: OpenAlex, cross-checked against Web of Science.
- Disciplines: biology, medicine, chemistry, physics, materials science and geology.
- AI eras: machine learning (1980–2014), deep learning (2015–2022) and generative AI (2023 onward).
- Detection: a BERT language model identified 310,957 papers (0.75% of the total) as AI-augmented, with an F1-score of 0.875 against expert labels.
- Researchers: 5,377,346 AI-adopting researchers were identified; career analysis covered 2,282,029 scientists.
The Individual Benefits
- Output: AI-adopting scientists publish 3.02 times as many papers.
- Citations: AI adopters receive 4.84 times as many citations.
- Paper-level impact: annual citations to AI papers are 98.70% higher than to non-AI papers.
- Career speed: AI adopters become research project leaders 1.37 years earlier, at 7.33 years compared with 8.70 years.
- Junior scientists: junior researchers who use AI have a 45.00% probability of becoming established researchers, 13.64% higher than juniors who do not.
The Collective Costs
- Smaller teams: the number of junior scientists per team falls from 2.89 to 1.99 in AI-using teams, a 31.14% drop.
- Narrower topics: AI adoption reduces the range of topics studied (“knowledge extent”) by a median 4.63%.
- Breadth of the effect: the contraction appears in all six disciplines and in more than 70% of over 200 subfields.
- Less engagement: AI research shows 22.00% less follow-on engagement; papers that cite the same work cite each other less.
- Citation concentration: 22.20% of top AI papers receive 80% of citations, and the Gini coefficient is 0.754 for AI papers compared with 0.690 for non-AI papers.
The authors call these popular but weakly interconnected research areas “lonely crowds”.
Why Science Narrows
- AI users move toward data-rich areas where AI performs well on benchmarks.
- Data-poor areas, where AI tools are less effective, receive less attention.
- The authors recommend rewarding research in data-poor fields and building AI for exploration rather than optimisation.
How Researchers Use AI
- Wiley ExplanAItions 2025: among 2,430 researchers, AI use rose from 57% in 2024 to 84% in 2025.
- Research tasks: use of AI for research and publication tasks rose from 45% to 62%.
- Concerns: worry about inaccuracies and hallucinations rose from 51% to 64%.
- Elsevier (November 2025): 58% of more than 3,200 researchers in 113 countries use AI tools, up from 37% in 2024.
- Trust: 22% of Elsevier respondents consider AI tools trustworthy, and 27% say they have adequate AI training.
- Pressure: 68% say pressure to publish has grown over the past two to three years.
AI and Isolation at Work
- Journal of Applied Psychology (2023): a study led by Pok Man Tang of the University of Georgia surveyed workers in the U.S., Taiwan, Indonesia and Malaysia. Frequent work with AI was linked to more loneliness and insomnia, as well as more helping behaviour toward coworkers.
- “The Cybernetic Teammate” (2025): in a field experiment with 776 Procter & Gamble professionals, individuals working with AI matched the performance of two-person teams without AI.
- Emotional effect: participants working with AI reported more positive emotions, which the authors suggest partly replaces the social role of human teammates.
People also form social bonds with machines outside work. A 2024 survey by the University of Waterloo and University College London found that two-thirds of 300 U.S. adults attribute some consciousness to ChatGPT, a pattern explored in I Asked My Robot Lamp How Its Day Was.
Research Networks Online
Research consortia, open-data repositories and academic collaboration networks often publish on the .net extension, which was originally intended for network organisations. A dedicated .net domain is commonly used for shared datasets, preprint servers and multi-lab project sites.
Key Numbers at a Glance
- 41.3 million: papers analysed in the Nature study.
- 3.02×: papers published by AI-adopting scientists.
- 4.84×: citations received by AI adopters.
- 4.63%: median reduction in topic range.
- 22%: drop in follow-on engagement in AI research.









