Artificial Intelligence Tools Expand Scientists' Impact but Contract Science's Focus (Just accepted by Nature, to be online soon)

TL;DR

The study empirically quantifies how AI-augmented science affects individual scientists and the collective landscape using 41.3 million OpenAlex papers across six natural science fields. A two-stage BERT-based method identifies AI-augmented papers with high accuracy, revealing that AI adopters achieve substantially higher productivity and citations, and reach established status earlier. However, AI-augmented research concentrates on a narrower set of topics, reduces follow-on engagement, and contracts the overall knowledge extent, suggesting a tension between personal advancement and collective scientific breadth. Robustness tests across the generative AI era and independent WoS data corroborate these findings, with data abundance emerging as a key driver of AI selectivity across topics and fields.

Abstract

Development in Artificial Intelligence (AI) has accelerated scientific discovery. Alongside recent AI-oriented Nobel prizes, these trends establish the role of AI tools in science. This advancement raises questions about the potential influences of AI tools on scientists and science as a whole, and highlights a potential conflict between individual and collective benefits. To evaluate, we used a pretrained language model to identify AI-augmented research, with an F1-score of 0.875 in validation against expert-labeled data. Using a dataset of 41.3 million research papers across natural science 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 scientist's engagement with one another by 22.00%. Thereby, AI adoption in science presents a seeming paradox -- an expansion of individual scientists' impact but a contraction in collective science's reach -- as AI-augmented work moves collectively toward areas richest in data. With reduced follow-on engagement, AI tools appear to automate established fields rather than explore new ones, highlighting a tension between personal advancement and collective scientific progress.

Figures (51)

• Figure 1: Increasing prevalence of AI adoption in science.(a) Increasing performance of AI paper identification during the two-stage fine-tuning of the BERT pre-trained models, where we use rough training data in Stage 1 to evolve precise assessments in Stage 2. We independently train two models based on titles and abstracts, respectively, and then integrate them into an ensemble that selects the optimal models during both stages (red stars) to identify all selected papers. (b) Accuracy evaluation of our identification results by human experts. For samples spanning three eras of AI, experts reached consensus with Fleiss’ Kappa ($\kappa$) $\geq 0.93$. Our model identification results have strong accuracy in validation against expert-labeled data with an F1-score $\geq 0.85$. (c) Relative adoption frequency of the top 15 AI methods across all disciplines for all selected AI development eras. (d)-(e) The growth of AI-augmented papers (d, $n=41,298,433$) and AI-adopted researchers (e, $n=5,377,346$) across the eras of ML, DL, and GAI between 1980 and 2025 in selected scientific disciplines. The y-axes are set to log-scale. (f) The average monthly growth rates for AI papers and researchers across the eras of ML, DL, and GAI across all selected disciplines ($n=543$ month observations), where 99% CIs are shown as error bars centred at the mean.
• Figure 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.
• Figure 1: Probability of AI (orange) and number of papers (green) for selected venues. Final identification results combine the best models in both stages of fine-tuning. Venues are ordered according to the probability of AI papers within them.
• Figure 2: AI enlarges paper impact and enhances researcher careers.(a) Average (insets: top 1% and 10%) annual citations after publication of AI and non-AI papers ($n=27,405,011$), where AI papers attract more citations. (b) Average annual citations for researchers adopting AI and their counterparts without AI ($p<0.001, n=5,377,346$), where researchers adopting AI garner 4.84 times more citations than their counterparts without AI. (c) The probability of two role transitions between junior scientists adopting AI and their counterparts without AI ($n=46$ year observations for each field). Junior scientists adopting AI have a higher probability of becoming established researchers and a lower probability of exiting academia compared with their counterparts without AI. (d) Survival functions for the transition from junior to established researcher ($p<0.001, n=2,282,029$). The survival function can be well-fit with exponential distributions, where junior scientists adopting AI become established earlier than their counterparts without AI. For all panels, 99% CIs are shown as error bars, with the insets of panel (a) centred at the 1% and 10% percentiles and other panels centred at the mean. All statistical tests use a two-sided t-test.
• Figure 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.