Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

  • Article
  • Published:

Artificial intelligence tools expand scientists’ impact but contract science’s focus

Nature volume 649, pages 1237–1243 (2026)Cite this article

Subjects

Abstract

Developments in artificial intelligence (AI) have accelerated scientific discovery1. Alongside recent AI-oriented Nobel prizes2,3,4,5,6,7,8,9, these trends establish the role of AI tools in science10. This advancement raises questions about the influence of AI tools on scientists and science as a whole, and highlights a potential conflict between individual and collective benefits11. To evaluate these questions, we used a pretrained language model to identify AI-augmented research, with an F1-score of 0.875 in validation against expert-labelled data. Using a dataset of 41.3 million research papers across the natural sciences and covering distinct eras of AI, here we show an accelerated adoption of AI tools among scientists and consistent professional advantages associated with AI usage, but a collective narrowing of scientific focus. Scientists who engage in AI-augmented research publish 3.02 times more papers, receive 4.84 times more citations and become research project leaders 1.37 years earlier than those who do not. By contrast, AI adoption shrinks the collective volume of scientific topics studied by 4.63% and decreases scientists’ engagement with one another by 22%. By consequence, adoption of AI in science presents what seems to be a paradox: an expansion of individual scientists’ impact but a contraction in collective science’s reach, as AI-augmented work moves collectively towards areas richest in data. With reduced follow-on engagement, AI tools seem to automate established fields rather than explore new ones, highlighting a tension between personal advancement and collective scientific progress.

This is a preview of subscription content, access via your institution

Access options

Buy this article

  • Purchase on SpringerLink
  • Instant access to the full article PDF.

USD 39.95

Prices may be subject to local taxes which are calculated during checkout

Fig. 1: Increasing prevalence of AI adoption in science.
Fig. 2: AI enlarges paper impact and enhances researcher careers.
Fig. 3: AI adoption is associated with a contraction in knowledge extent within and across scientific fields.
Fig. 4: Reduced follow-on engagement and more overlapping works in AI research.

Similar content being viewed by others

Data availability

The OpenAlex dataset for research papers and researchers is available at https://docs.openalex.org/download-all-data/openalex-snapshot. The Web of Science dataset for research papers and researchers is available at https://clarivate.com/academia-government/scientific-and-academic-research/research-discovery-and-referencing/web-of-science/web-of-science-core-collection. The Journal Citation Report dataset for the journal quantile is retrieved from https://jcr.clarivate.com/jcr/browse-journals. The author contribution dataset is available at https://zenodo.org/records/6569339. The pre-trained parameters for the BERT language model are available at https://huggingface.co/docs/transformers. The pre-trained parameters for the SPECTER 2.0 text embedding model are available at https://huggingface.co/allenai/specter2. Source data are provided with this paper.

Code availability

This study used Python 3.11.0 with software packages to conduct data analysis. Required packages are NumPy (v.1.26.4), pandas (v.2.2.3), SciPy (v.1.15.2), scikit-learn (v.1.6.1) and matplotlib (v.3.10.1). The t-SNE algorithm used is imported from the sklearn package. The codes developed in this study can be found at https://github.com/tsinghua-fib-lab/AI-Impacts-Science.

References

  1. Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47–60 (2023).

    ADS  CAS  PubMed  Google Scholar 

  2. Hopfield, J. J. Neural networks and physical systems with emergent collective computational abilities. Proc. Natl Acad. Sci. USA 79, 2554–2558 (1982).

    ADS  MathSciNet  CAS  PubMed  PubMed Central  Google Scholar 

  3. Hopfield, J. J. Neurons with graded response have collective computational properties like those of two-state neurons. Proc. Natl Acad. Sci. USA 81, 3088–3092 (1984).

    ADS  CAS  PubMed  PubMed Central  Google Scholar 

  4. LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015).

    ADS  CAS  PubMed  Google Scholar 

  5. Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. Commun. ACM, 60, 84–90 (2012).

  6. Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science 313, 504–507 (2006).

    ADS  MathSciNet  CAS  PubMed  Google Scholar 

  7. Hinton, G. E. Training products of experts by minimizing contrastive divergence. Neural Comput. 14, 1771–1800 (2002).

    PubMed  Google Scholar 

  8. Kuhlman, B. et al. Design of a novel globular protein fold with atomic-level accuracy. Science 302, 1364–1368 (2003).

    ADS  CAS  PubMed  Google Scholar 

  9. Jumper, J. et al. Highly accurate protein structure prediction with alphafold. Nature 596, 583–589 (2021).

    ADS  CAS  PubMed  PubMed Central  Google Scholar 

  10. Gao, J. & Wang, D. Quantifying the use and potential benefits of artificial intelligence in scientific research. Nat. Human Behav. 8, 2281–2292 (2024).

  11. Evans, J. A. Electronic publication and the narrowing of science and scholarship. Science 321, 395–399 (2008).

    ADS  CAS  PubMed  Google Scholar 

  12. Adıgüzel, T., Kaya, M. H. & Cansu, F. K. Revolutionizing education with AI: exploring the transformative potential of ChatGPT. Contemp. Educat. Technol. 15, ep429 (2023).

  13. Akgun, S. & Greenhow, C. Artificial intelligence in education: addressing ethical challenges in K-12 settings. AI Ethics 2, 431–440 (2022).

    PubMed  Google Scholar 

  14. Meskó, B. & Topol, E. J. The imperative for regulatory oversight of large language models (or generative AI) in healthcare. npj Digital Med. 6, 120 (2023).

    Google Scholar 

  15. Loh, H. W. et al. Application of explainable artificial intelligence for healthcare: a systematic review of the last decade (2011–2022). Comput. Methods Prog. Biomed. 226, 107161 (2022).

  16. Ahmed, I., Jeon, G. & Piccialli, F. From artificial intelligence to explainable artificial intelligence in industry 4.0: a survey on what, how, and where. IEEE Trans. Indust. Inform. 18, 5031–5042 (2022).

    Google Scholar 

  17. Varadi, M. et al. Alphafold protein structure database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucl. Acids Res. 50, D439–D444 (2022).

    CAS  PubMed  Google Scholar 

  18. Degrave, J. et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature 602, 414–419 (2022).

    ADS  CAS  PubMed  PubMed Central  Google Scholar 

  19. Fawzi, A. et al. Discovering faster matrix multiplication algorithms with reinforcement learning. Nature 610, 47–53 (2022).

    ADS  CAS  PubMed  PubMed Central  Google Scholar 

  20. Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570–578 (2023).

    ADS  CAS  PubMed  PubMed Central  Google Scholar 

  21. Stokel-Walker, C. & Van Noorden, R. What ChatGPT and generative AI mean for science. Nature 614, 214–216 (2023).

    ADS  CAS  PubMed  Google Scholar 

  22. Gilson, A. et al. How does ChatGPT perform on the United States medical licensing examination? The implications of large language models for medical education and knowledge assessment. JMIR Med. Educat. 9, e45312 (2023).

    Google Scholar 

  23. Salimi, A. & Saheb, H. Large language models in ophthalmology scientific writing: ethical considerations blurred lines or not at all? Am. J. Ophthalmol. 254, 177–181 (2023).

    PubMed  Google Scholar 

  24. Liang, W. et al. Mapping the increasing use of LLMs in scientific papers. In Proc. 1st Conference on Language Modeling (COLM, USA, 2024).

  25. Hwang, T. et al. Can ChatGPT assist authors with abstract writing in medical journals? Evaluating the quality of scientific abstracts generated by ChatGPT and original abstracts. PLoS ONE 19, e0297701 (2024).

    CAS  PubMed  PubMed Central  Google Scholar 

  26. Kobak, D., González-Márquez, R., Horvát, E.-Á. & Lause, J. Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Sci. Adv. 11, eadt3813 (2025).

    ADS  PubMed  PubMed Central  Google Scholar 

  27. Wojtowicz, Z. & DeDeo, S. Undermining Mental Proof: How AI Can Make Cooperation Harder by Making Thinking Easier Vol. 39, 1592–1600 (2025).

  28. Frank, M. R., Wang, D., Cebrian, M. & Rahwan, I. The evolution of citation graphs in artificial intelligence research. Nat. Mach. Intell. 1, 79–85 (2019).

    Google Scholar 

  29. OpenAlex (OpenAlex, 2025); https://openalex.org/.

  30. Clarivate (Web of Science, 2025); https://www.webofscience.com.

  31. Mongeon, P. & Paul-Hus, A. The journal coverage of web of science and scopus: a comparative analysis. Scientometrics 106, 213–228 (2016).

    Google Scholar 

  32. Devlin, J. et al. BERT: pre-training of deep bidirectional transformers for language understanding. In Proc. 57th Annual Meeting of the Association for Computational Linguistics 4171–4186 (ACL, Italy, 2019).

  33. Wolf, T. et al. Transformers: State-of-the-art natural language processing. In Proc. 58th Annual Meeting of the Association for Computational Linguistics 38–45 (ACL, 2020).

  34. Beltagy, I., Lo, K. & Cohan, A. SciBERT: a pretrained language model for scientific text. In Proc. 57th Annual Meeting of the Association for Computational Linguistics 3613–3618 (ACL, Italy, 2019).

  35. Cohan, A., Feldman, S., Beltagy, I., Downey, D. & Weld, D. S. SPECTER: document-level representation learning using citation-informed transformers. In Proc. 58th Annual Meeting of the Association for Computational Linguistics 2270–2282 (ACL, 2020).

  36. Singh, A., D’Arcy, M., Cohan, A., Downey, D. & Feldman, S. SciRepEval: a multi-format benchmark for scientific document representations. In Proc. 61st Annual Meeting of the Association for Computational Linguistics 5548–5566 (ACL, Canada, 2023).

  37. Landis, J. R. & Koch, G. G. The measurement of observer agreement for categorical data. Biometrics 33, 159–174 (1977).

  38. Fleiss, J. L. Measuring nominal scale agreement among many raters. Psychol. Bull. 76, 378 (1971).

    Google Scholar 

  39. Chu, J. S. & Evans, J. A. Slowed canonical progress in large fields of science. Proc. Natl Acad. Sci. USA 118, e2021636118 (2021).

    CAS  PubMed  PubMed Central  Google Scholar 

  40. Journal Citation Reports (Clarivate, 2021); https://jcr.clarivate.com/jcr/home.

  41. Ioannidis, J. P., Boyack, K. W. & Klavans, R. Estimates of the continuously publishing core in the scientific workforce. PloS ONE 9, e101698 (2014).

    ADS  PubMed  PubMed Central  Google Scholar 

  42. Kendall, D. G. Birth-and-death processes, and the theory of carcinogenesis. Biometrika 47, 13–21 (1960).

    MathSciNet  Google Scholar 

  43. Fortunato, S. et al. Science of science. Science 359, eaao0185 (2018).

    PubMed  PubMed Central  Google Scholar 

  44. Milojević, S. Quantifying the cognitive extent of science. J. Informetrics 9, 962–973 (2015).

    Google Scholar 

  45. McMahan, P. & Evans, J. Ambiguity and engagement. Am. J. Sociol. 124, 860–912 (2018).

    Google Scholar 

  46. Merton, R. K. The matthew effect in science: the reward and communication systems of science are considered. Science 159, 56–63 (1968).

    ADS  CAS  PubMed  Google Scholar 

  47. Borger, J. G. et al. Artificial intelligence takes center stage: exploring the capabilities and implications of chatgpt and other AI-assisted technologies in scientific research and education. Immunol. Cell Biol. 101, 923–935 (2023).

    PubMed  Google Scholar 

  48. Lawrence, N. D. & Montgomery, J. Accelerating AI for science: open data science for science. Royal Soc. Open Sci. 11, 231130 (2024).

    ADS  Google Scholar 

  49. King, R. D. et al. The automation of science. Science 324, 85–89 (2009).

    ADS  CAS  PubMed  Google Scholar 

  50. Burger, B. et al. A mobile robotic chemist. Nature 583, 237–241 (2020).

    ADS  CAS  PubMed  Google Scholar 

  51. Krauss, A. Debunking Revolutionary Paradigm Shifts: Evidence of Cumulative Scientific Progress Across Science Vol. 480, 20240141 (The Royal Society, 2024).

  52. Microsoft Academic Graph (Microsoft, 2015); https://www.microsoft.com/en-us/research/project/microsoft-academic-graph.

  53. Open Academic Graph (Aminer, 2020); https://old.aminer.cn/oag-2-1.

  54. Porter, A. & Rafols, I. Is science becoming more interdisciplinary? Measuring and mapping six research fields over time. Scientometrics 81, 719–745 (2009).

    ADS  Google Scholar 

  55. Rumelhart, D. E., Hinton, G. E. & Williams, R. J. Learning representations by back-propagating errors. Nature 323, 533–536 (1986).

    ADS  Google Scholar 

  56. LeCun, Y. et al. Backpropagation applied to handwritten zip code recognition. Neural Comput. 1, 541–551 (1989).

    Google Scholar 

  57. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual Learning for image recognition. In CVPR'16: Proc. 2016 IEEE conference on computer vision and pattern recognition 770–778 (2016).

  58. Face, H. Bert for Sequence Classification (Hugging Face, 2025); https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertForSequenceClassification.

  59. Sekara, V. et al. The chaperone effect in scientific publishing. Proc. Natl Acad. Sci. USA 115, 12603–12607 (2018).

    ADS  CAS  PubMed  PubMed Central  Google Scholar 

  60. Chen, T. & Guestrin, C. XGBoost: a scalable tree boosting system. In KDD'16: Proc. 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 785–794 (2016).

  61. Hill, R. et al. The pivot penalty in research. Nature 642, 999–1006 (2025).

  62. Milojević, S., Radicchi, F. & Walsh, J. P. Changing demographics of scientific careers: the rise of the temporary workforce. Proc. Natl Acad. Sci. USA 115, 12616–12623 (2018).

    ADS  PubMed  PubMed Central  Google Scholar 

  63. Xu, F., Wu, L. & Evans, J. Flat teams drive scientific innovation. Proc. Natl Acad. Sci. USA 119, e2200927119 (2022).

    CAS  PubMed  PubMed Central  Google Scholar 

  64. Lin, Y., Frey, C. B. & Wu, L. Remote collaboration fuses fewer breakthrough ideas. Nature 623, 987–991 (2023).

    ADS  CAS  PubMed  Google Scholar 

  65. Kingman, J. F. C. Poisson Processes Vol. 3 (Clarendon, 1992).

  66. Meisling, T. Discrete-time queuing theory. Operat. Res. 6, 96–105 (1958).

    MathSciNet  Google Scholar 

Download references

Acknowledgements

This work was supported in part by the National Natural Science Foundation of China (grant no. U23B2030, 23IAA02114 and 62472241), the joint project of Infinigence AI & Tsinghua University, and Tsinghua University-Toyota Research Institute to Y. L. and F.X. J.E. received support from Novo Nordisk Foundation (Simulations of Science for Society), NSF (grant no. 2404109) and the United States Department of Defense (Defense Advanced Research Projects Agency - Modeling and Measuring Scientific Creativity). The funders had no role in study design, data collection, analysis, preparation of or decision to publish the manuscript.

Author information

Authors and Affiliations

  1. Department of Electronic Engineering, Beijing National Research Center for Information Science and Technology (BNRist), Tsinghua University, Beijing, P. R. China

    Qianyue Hao, Fengli Xu & Yong Li

  2. Zhongguancun Academy, Beijing, P. R. China

    Yong Li

  3. Knowledge Lab and Department of Sociology, University of Chicago, Chicago, IL, USA

    James Evans

  4. Santa Fe Institute, Santa Fe, NM, USA

    James Evans

Authors
  1. Qianyue Hao
    View author publications

    Search author on:PubMed Google Scholar

  2. Fengli Xu
    View author publications

    Search author on:PubMed Google Scholar

  3. Yong Li
    View author publications

    Search author on:PubMed Google Scholar

  4. James Evans
    View author publications

    Search author on:PubMed Google Scholar

Contributions

F.X., Y.L. and J.E. jointly launched this research and designed the research outline. Q.H. analysed the data and prepared the figures. All authors jointly participated in writing and revising the manuscript.

Corresponding authors

Correspondence to Fengli Xu, Yong Li or James Evans.

Ethics declarations

Competing interests

J.E. has a commercial affiliation with Google, but Google had no role in the design, analysis, or decision to publish this study. The authors declare no other competing interests.

Peer review

Peer review information

Nature thanks Giovanni Colavizza, Luis Nunes Amaral, Catherine Shea, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended Data Fig. 1 Illustration for the method of identifying AI usage in research papers with fine-tuned language models.

(a) Structure of our deployed language model, which consists of the tokenizer, the core BERT model, and the linear layer. (b) Procedure of the two-stage model fine-tuning process, where we design specific approaches for constructing positive and negative data at each stage.

Extended Data Fig. 2 Procedure of accuracy evaluation via expert evaluation.

We randomly sample 1320 papers and delegate three experts to scrutinize the identification results for each paper. We then draw the final expert label of each paper from the three experts according to the principle of the minority obeying the majority and validate the result of the language model with it. Results indicate strong consistency among experts and high accuracy with our identification results.

Extended Data Fig. 3 Comparison of the total citations of AI and non-AI papers published in different eras.

Results show that AI papers consistently attract more citations over different eras (P < 0.001, n = 27,405,011), indicating a higher academic impact than non-AI papers. 99% CIs are shown as error bars centred at the mean, and the statistical tests use a two-sided t-test.

Source data

Extended Data Fig. 4 Annual publications of researchers adopting AI and their counterparts without AI.

Results show that in all 6 scientific disciplines, researchers adopting AI are more productive than their counterparts without AI (P < 0.001, n = 5,377,346). On average, researchers adopting AI annually publish 3.02 times more papers compared with those not using AI. 99% CIs are shown as error bars centred at the mean, and the statistical tests use a two-sided t-test.

Source data

Extended Data Fig. 5 Scientists’ career role transition.

(a) The career role transition of researchers. We consider the last author of each paper as research project leader and researchers who have been research project leaders as established researchers. Researchers who have yet to lead a research project are junior researchers, and they have two potential role transition pathways in the future: (1) become established researchers (solid arrow), and (2) abandon academia (dashed arrow). (b) Change in the ratio of conceptual work across the research career, before and after becoming an established researcher. The ratio increases rapidly before the role transition to established researchers, while it remains stable and high after that transition. 99% CIs are shown as error bands centred at the mean.

Source data

Extended Data Fig. 6 Team composition of AI and non-AI papers.

(a) AI research is associated with reduced research team sizes, averaging 1.33 fewer scientists (P < 0.001, n = 33,528,469). Specifically, the average number of junior scientists decreased from 2.89 in non-AI teams to 1.99 in AI teams (31.14%), while the number of established scientists decreased from 4.01 to 3.58 (10.77%). (b)-(d) Change in team size, average number of junior researchers, and average number of established researchers. These findings indicate that within the overall trend of increasing size of scientific research teams, AI adoption primarily contributes to a reduction in the number of junior scientists in teams, while a decrease in the number of established scientists is more moderate. (e) The average career age of team leaders in AI and non-AI papers. (f) The average career age of all involved established researchers in AI and non-AI papers. Results indicate that AI accelerates the transition from junior to established scientists, enabling AI-adopted researchers to become established at a younger age than those without AI. For all panels, 99% CIs are shown as error bars or error bands centred at the mean. All statistical tests use a two-sided t-test.

Source data

Extended Data Fig. 7 Model fitting the role transition time of junior scientists.

(a) (c) (e) Survival functions for the transition from junior to established researcher in (a) biology (n = 625,093), (c) medicine (n = 1,137,076), and (e) physics (n = 120,366). (b) (d) (f) Survival functions for the transition from junior researcher to leave academia in (b) biology (n = 625,093), (d) medicine (n = 1,137,076), and (f) physics (n = 120,366). All survival functions can be well-fit with exponential distributions, where the expected time for junior scientists to become established is shorter for those who adopt AI (P < 0.001), while the expected time for junior scientists to abandon academia is similar or slightly longer for those who adopt AI. Results indicate that AI not only provides junior scientists opportunities to become established scientists at a younger age, but also reduces the risk of their exiting academia early. For all panels, 99% CIs are shown as error bars centred at the mean. All statistical tests use a two-sided t-test.

Source data

Extended Data Fig. 8 The knowledge extent of AI and non-AI papers.

Here we visualize the embeddings of a small random sample of 2,000 papers, half of which are AI papers and half are non-AI papers. To eliminate randomness introduced by the t-SNE algorithm, here we simply pick out the first two dimensions of the high-dimensional embeddings to flatten them into a 2-D plot, and we provide 5 different random batches for each field to ensure robustness. As shown by the solid arrows and circular boundaries, the knowledge extent of AI papers is smaller than that of a comparable sample of non-AI papers, which is consistent across the fields studied in our analysis.

Extended Data Fig. 9 The knowledge extent of AI and non-AI papers in each subfield.

Compared with conventional research, AI research is associated with a shrinkage in the collective knowledge extent of science, where the contraction of knowledge extent can be observed in more than 70% of over two hundred sub-fields (n = 1,000 samples in each subfield). For all subfields, 99% CIs are shown as error bars centred at the mean.

Source data

Extended Data Fig. 10 The Matthew effect in citations to AI and non-AI papers.

In AI research, a small number of superstar papers dominate the field, with approximately 20% of top papers receiving 80% of citations and 50% receiving 95%. This unequal distribution leads to a higher Gini coefficient in citation patterns surrounding AI research (P < 0.001, n = 100 sampled paper groups for each discipline). Such disparity in the recognition of AI papers is consistent across all fields examined. For all panels, 99% CIs are shown as error bars or error bands centred at the mean. All statistical tests use a two-sided t-test.

Source data

Supplementary information

Supplementary Information

Supplementary Sections 1–4, Figs. 1–37 and Tables 1–12. These provide further detail and background information, and numerous extended analyses and robustness tests to the main results.

Reporting Summary

Peer Review file

Supplementary Data

Source data for Supplementary Figs. 1 and 4–37.

Source data

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Hao, Q., Xu, F., Li, Y. et al. Artificial intelligence tools expand scientists’ impact but contract science’s focus. Nature 649, 1237–1243 (2026). https://doi.org/10.1038/s41586-025-09922-y

Download citation

  • Received:

  • Accepted:

  • Published:

  • Version of record:

  • Issue date:

  • DOI: https://doi.org/10.1038/s41586-025-09922-y

Search

Advanced search

Quick links

Nature Briefing Careers

Sign up for the Nature Briefing: Careers newsletter — what matters in careers research, free to your inbox weekly.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing: Careers