Introduction

Artificial intelligence (AI) promises to revolutionize science and technology, sparking widespread discussions about its implications, including for ethics (OECD, 2023), reproducibility (Haibe-Kains et al., 2020), employment (Septiandri et al., 2024), productivity (Filippucci et al., 2024), and creativity (Mukherjee & Chang, 2024). These discussions have been heightened by the recent and rapid rise of Generative Artificial Intelligence (GenAI)—a type of AI that uses machine learning to dynamically generate new content from large amounts of training data (Lorenz et al., 2023; Stryker & Scapicchio, 2024). GenAI (which encompasses large language models such as ChatGPT) has garnered sizable attention within scientific communities. A Nature survey revealed substantial interest in generative AI tools among researchers, with a significant proportion seeking to use these technologies to brainstorm ideas and conduct research (Van Noorden & Perkel, 2023). Others have highlighted GenAI’s potential to accelerate the pace of science, boost the discovery of new drugs and materials, transform experimental design and hypothesis formulation, and disrupt scholarly publishing (Charness et al., 2023; Harries et al., 2025; Liu et al., 2023; Lund et al., 2023; Vert, 2023).

However, notwithstanding its rapid spread, the deployment of generative AI remains fraught with ethical and epistemic challenges (Blau et al., 2024; Charness et al., 2023). Moreover, the non-transparent nature of proprietary generative AI systems, exemplified by ChatGPT’s closed-source architecture, raises fundamental questions regarding intellectual property rights and ethical responsibilities in scientific research (Liverpool, 2023; Lund et al., 2023). Alongside mounting debate in research communities about the promises and pitfalls of GenAI, new landscape and bibliometric studies are being produced which seek to assess the growth, scale and scope of GenAI in science (for example, see: Dwivedi & Elluri, 2024; Pathak & Pallasena, 2025). These extant studies typically identify relevant publications using a bibliometric search for GenAI; the results are then analyzed, including by author, institutional, subject, and geographical aspects, without determining how GenAI is used in practice in these publications. This distinction is vital given the highly visible debate about the implications of GenAI in science (and academia in general) that has arisen alongside the remarkably rapid emergence and ease of accessibility to GenAI models.

In this study, we offer an advance which distinguishes between papers that apply GenAI in their research and those that study GenAI as a research object. Conceptually and methodologically, this is a critical distinction as it allows us to probe whether researchers are actively incorporating GenAI into their methods and workflows, or whether they are primarily considering GenAI’s implications, potentials, or risks but without applying it. This distinction is also relevant for AI research across the board, where it sheds light on whether and how researchers, across all disciplines and countries, are embracing AI (including but not limited to GenAI) as a valuable aid to science and the extent to which they are discussing its implications.

With this distinction in mind, we investigate three critical research questions to better understand the diffusion and potential impact of GenAI in science. First, we explore how and where GenAI is being used in science. Recent studies have observed the rapid adoption of GenAI tools in various scientific disciplines, reflecting their growing influence on research methodologies and knowledge production (Jin et al., 2023; Xiaoyu et al., 2025). As indicated, we take the further step of distinguishing between applications of GenAI and discussions of GenAI but without actual implementation, with comparison to other AI technologies. Understanding such patterns in the use of GenAI across different scientific fields and nations can provide insights into its role in shaping contemporary scientific practices.

Second, we investigate how GenAI relates to collaboration patterns in scientific research, again with comparison to other AI technologies. It is reported that GenAI is helping some researchers to write code, review literature, and prepare presentations and manuscripts (Owens, 2023), with Noy and Zhang (2023) finding in a controlled experiment that college-educated professionals using ChatGPT saw substantial productivity gains across their various writing tasks. Conceivably, if GenAI leads to increased productivity across the range of tasks undertaken by researchers, this might lead to reduced team sizes for research outputs that focus on applying GenAI. Indeed, concerns have been raised about the potential of GenAI to automate tasks and roles, with job impacts for human researchers (Kim, 2023). On the other hand, as scientific research becomes increasingly specialized and collaborative (Jones, 2009), with large teams often required to tackle complex problems across various disciplines (Venturini et al., 2024), it is also conceivable that GenAI might lead to expanded team sizes. In short, the link between GenAI use and team size or composition is still unclear. Author team size is also associated with the performance of research outcomes. Prior findings underscore that large teams tend to advance established knowledge, whereas smaller teams are more likely to drive scientific and technological disruption (Wu et al., 2019). Understanding whether GenAI might reduce the need for large, diverse teams, or conversely, necessitate even larger collaborations, is important for anticipating the disruptiveness of the research outputs.

Our third question concerns the association between GenAI adoption and international collaboration. International collaboration is often correlated with high-quality research outcomes and increased citation rates (Wang et al., 2024). However, GenAI is growing in a period of heightened geopolitical conflict and concern about research security involving China, the United States, Russia, Europe, and other advanced research nations, posing new challenges to global scientific collaboration (Jia et al., 2024; Min et al., 2023; Shih et al., 2024). In this context, examining the correlation between GenAI adoption and international collaboration compared to other AI technologies is particularly timely. The intersection of GenAI application and discussion modes of research with that of international collaboration offers a rich avenue for understanding the broader implications of GenAI in a rapidly changing global landscape.

To address these questions, this paper presents an exploratory bibliometric analysis of the rise of GenAI in scientific research. Using OpenAlex, which provides comprehensive scientific publication metadata (Priem et al., 2022), we analyze selected populations of publications to investigate the characteristics of GenAI compared to other AI technologies. In this respect, we seek to explore the additional consequences of GenAI compared with developments already underway through the deployment of earlier and more established AI technologies. Given the limitation of bibliometric searches, which can retrieve relevant publications without indicating whether AI is actually being applied, we use a two-stage classification approach based on GPT and SciBERT (Beltagy et al., 2019). In addition to identifying GenAI and other AI publications, we also use this approach to categorize papers into application publications that focus on applying AI technologies and discussion publications that primarily discuss AI without implementing it. Furthermore, we examine the disciplinary and geographic diffusion and use of GenAI, as well as its relationship with team size and international collaboration. The findings on these topics contribute to the ongoing discourse on GenAI’s role in science, providing insights that are relevant for research initiatives and science policy.

Data and methodology

To investigate the rise of GenAI in science, our analysis draws on a dataset of AI-related publications (referred to as All AI Publications), which includes both GenAI-related and other AI-related works. This dataset is further categorized into an AI Application Publications sub-dataset and an AI Discussion Publications sub-dataset. Further details are provided in following sections.

Bibliometric data collection

We construct a bibliometric sample containing both GenAI papers and other artificial intelligence (Other-AI) papers to investigate the diffusion of GenAI in science compared with other AI technologies. We employ OpenAlex as our primary data source, leveraging its extensive repository of over 250 million scientific works, encompassing journal articles, conference proceedings, and more. The dataset is sourced from the January 2024 snapshot of OpenAlex, and our study retains papers published before 2024. We use a two-step process to get an initial set of All AI Publications.

In the first step, to identify GenAI related papers in OpenAlex, we build a list of search keywords, informed by prior bibliometric studies. Kanbach et al. (2024) composed a bibliometric definition for generative AI. Although there are multiple GenAI models, they demonstrated that there is a clear tendency for most of these publications to reference ChatGPT as the most prominent use of GenAI. In contrast, Mariani and Dwivedi (2024) deployed a list of model-name keywords to cover GenAI, including generative AI models such as “Runway” and “Bard”. However, we find that these and other additional keywords add noise into the retrieved results where GenAI models are named after commonly found terms that are not unique to GenAI.Footnote 1 We thus proceeded by formulating a list of terms and phrases specific to generative AI that includes search terms commonly used in e.g., China such as Artificial Intelligence Generated Content (AIGC), and with ChatGPT (which, at the time of writing, is by far the most widely and globally used GenAI model) serving as a representative model. From this list, we constructed PySpark SQL queries targeting the title and abstract fields. (See Table 1.) We further designed a Boolean search statement to filter SQL records based on specific string matches, employing the ‘LIKE’ operator along with wildcard characters ‘%’ to account for variations in terms. Both ‘LIKE’ and ‘ILIKE’ operators are utilized to differentiate between case-sensitive and case-insensitive terms.

Table 1 Generative AI search terms
Full size table

The publications returned by the search terms are not exclusive; for example, a publication containing both ‘chatgpt and ‘Generative AI’ would be captured by both search terms. We subsequently de-duplicate the records to address this. Additionally, we note that the acronym ‘GPT’ is also used elsewhere in science other than AI research. For example, GPT can refer to ‘Glutamate Pyruvate Transaminase’ in biotechnology research. To refine the dataset, we developed a preliminary set of GPT full-expansion terms through an intensive iterative process of manually checking the retrieved results to address the ambiguity of the abbreviation ‘GPT’. Table 2 presents the complete list of expansions for the acronym GPT that are excluded from our study. We further exclude irrelevant terms using the Boolean operator ‘NOT’. In addition, we restrict the time period to 2017–2023 to reduce noise. We choose 2017 as the starting point because it marks the release of the seminal Transformer model (Vaswani et al., 2017), which laid the foundation for modern generative AI. This model catalyzed a new wave of advancements in generative techniques, making 2017 a natural starting point for analyzing the diffusion of GenAI in scientific research. The endpoint of 2023 reflects the most recent full year of publication data available in the January 2024 release of the OpenAlex database, which we used for our analysis.

Table 2 Expansions of the acronym GPT
Full size table

In the second step, for Other-AI papers, we also used a keyword approach, beginning first with an approach to identify AI. We recognize that with its diverse interdisciplinary roots, varied development paths, and rapid recent growth (introducing new models and terms), ‘artificial intelligence’ is a challenging concept to define and quantify (Baruffaldi et al., 2020). Prior studies have constructed AI bibliometric search terms (Liu et al., 2021), although no consensus has been reached in the bibliometric community on defining AI. Considering the ongoing development of AI, we followed the AI search terms from Van Noorden and Perkel (2023), which was recent to our search, with the exclusion of ‘chatgpt’ in the terms. Specifically, we searched for papers in OpenAlex spanning 2017–2023, with titles or abstracts containing the following search terms: ‘machine learning’; ‘neural net*’, ‘deep learning’, ‘random forest’, ‘support vector machine’, ‘artificial intelligence’, ‘dimensionality reduction’, ‘gaussian process’, ‘naïve bayes’, ‘large language model’, ‘gaussian mixture model’, ‘ensemble methods’, and ‘llm’.Footnote 2

To check the accuracy of our search strategy, we assessed precision and recall following the methodology adopted by Schmallenbach et al. (2024). To assess recall, we examined the coverage of our search strategy in relevant conference proceedings or special issues. For AI-related terms, we used the conference series “Proceedings of the AAAI Conference on Artificial Intelligence” as a case study to estimate recall. This conference is among the most prestigious in the field of artificial intelligence and features the highest number of AI publications among all conference proceedings in our dataset. According to the OpenAlex database, a total of 10,261 publications were published in this conference series between 2017 and 2023. Our AI and GenAI search strategy retrieved 3074 of these, resulting in an estimated recall of approximately 30%. We note, however, that our study does not aim to construct a comprehensive dataset of all AI-related publications; rather, we prioritize the precision of the search strategy over recall. For GenAI-specific search terms, no conference proceedings or special issues exclusively focused on Generative AI were published within our time frame.

To assess precision, we randomly sampled 75 GenAI-related publications and 75 Other-AI-related publications with abstracts exceeding 100 words. Two independent reviewers annotated each publication for AI relevance in the abstract. In cases of disagreement, a third reviewer adjudicated. The resulting precision was 97.3% (146/150) for all AI-related search terms and 96% (72/75) for GenAI-related search terms. Of the 150 publications, four (three GenAI and one Other-AI) were non-English and thus could not be properly assessed, accounting for the observed annotation errors.

To deal with duplications, missing metadata and non-research publication records in our retrieved results, we implemented a series of data preprocessing procedures, to arrive at a refined dataset for subsequent analysis. Detailed steps are listed in the Appendix. After data cleaning, we obtained 14,729 GenAI-related entries and 1,255,089 AI-related entries, with 8266 publications overlapping between the two. To ensure a clear distinction between datasets, we further excluded all GenAI publications identified in the first step from the AI set, resulting in a non-overlapping dataset of Other-AI papers.

Data preprocessing

To explore the disciplinary diffusion of GenAI, we associated each of these papers with their subject to explore the interdisciplinary diffusion of GenAI technology. We followed the methodology introduced by Klebel and Ross-Hellauer (2023) to leverage the concepts entity provided within OpenAlex as the subject. We used the 19 root-level concepts from OpenAlex, which comprise (in alphabetical order) art, biology, business, chemistry, computer science, economics, engineering, environmental science, geography, geology, history, materials science, mathematics, medicine, philosophy, physics, political science, psychology, and sociology. OpenAlex assigns a score to indicate the strength of association between a publication and a given concept. Concepts with a score of 0 were filtered out from our analysis.

To investigate the geographical diffusion of GenAI, we examined the authorship country distribution. This analysis was complicated by the high proportion of missing institutional metadata in OpenAlex. Zhang et al. (2024) identified a prevalent issue of missing institution data in more than 60% of journal articles within the OpenAlex database. In OpenAlex, a publication is associated with authorships, each potentially linked to one or more institutions, delineated by five primary fields: institution, institution name, Research Organization Registry (ROR), country code, and institution type. In our dataset, among the 14,729 GenAI publications obtained from OpenAlex, 8825 (60.0%) publications had missing institutions. Among the 1,246,823 Other-AI publications, 407,205 publications (32.7%) had missing institutions. The term “missing institution” refers to the absence of any of the metadata including institution ID, institution name and institution country code.

To supplement the incomplete institutional metadata associated with GenAI-related publications—particularly those disseminated as arXiv preprints, which frequently lack affiliation information—we implemented a two-stage approach designed to enhance data completeness and ensure consistency across records. In the first stage, we retrieved institutional metadata at the paper level from the OpenAlex webpages. This procedure enabled us to supplement institutional information for 2748 publications. In the second stage, we focused on arXiv preprints by extracting affiliation data directly from PDF documents. Following plain-text conversion, we identified segments containing author and institution information, which were subsequently processed using GPT-3.5 to extract institution names and associated countries. The prompt used is shown in Fig. 1. To ensure consistency with OpenAlex standards, we employed a normalization procedure based on a term-frequency inverse-document-frequency (TF-IDF) vector space model. Cosine similarity scores were computed between the extracted institution names and the official OpenAlex institution names. The most similar entry was selected as the normalized institutional affiliation. A further quality control step was introduced to assess the reliability of country-level metadata. If the country name extracted from the PDF could be confidently mapped to an ISO2 country code, it was retained. Otherwise, the normalized institution was accepted only if the cosine similarity score exceeded a threshold of 0.7. Through this method, institutional data for 3332 out of 3637 arXiv publications were successfully resolved. The final authorship dataset was constructed by integrating institutional information from three sources: the original OpenAlex metadata, the OpenAlex webpage-crawled records, and the arXiv-based extraction. As a result of this comprehensive effort, 11,964 out of 14,729 GenAI publications (81.2%) were associated with complete institutional metadata, enabling meaningful downstream analyses.

Fig. 1
Fig. 1
Full size image

Prompts used to resolve missing institution information. Note: Illustration of prompt used with GPT-3.5 to extract missing institution names and corresponding countries from institution-incomplete OpenAlex publication metadata. See text for explanation and details of subsequent processing

Data classification

A keyword-based search strategy may capture publications that exhibit different uses of the AI search term. Hence, the next step in our approach involved classifying each AI publication based on whether it applied AI methods or focused on discussing AI without implementing them. Publication abstracts were necessary to undertake this classification. However, 65,193 publications lacked abstracts and were excluded from our analysis because they did not provide sufficient information for classification. We refer to the resulting sample of AI publications with complete abstract information as All AI Publications, which comprised a total of 786,389 publications. This dataset contained 10,395 GenAI entries including 3804 journal articles, 3803 preprints and 137 conference proceedings, and 775,994 Other-AI entries including 449,068 journal articles, 25,064 preprints and 45,203 conference proceedings. The detailed distribution of sources by type for GenAI and Other-AI set is shown in Table 5 in the Appendix.

To distinguish publications that primarily discuss AI from those that actively apply AI, we developed a two-stage classification model based on GPT-4o and SciBERT to classify publications into two categories: Application and Discussion. A publication was classified as Application if it incorporated AI models into its research methodology. This includes (1) advancing new or existing AI models, algorithms, or training methods (i.e., methodological contributions), or (2) using existing AI models as tools—without modification—in a research context. A publication was classified as Discussion if it focused on analyzing, critiquing, or reflecting on AI without implementing AI models in its methodology. This category includes literature reviews, ethical or societal commentary, perception studies, interviews, qualitative research, and bibliometric or meta-analyses. We acknowledge that application publications will typically include a section devoted to discussion and implications of the research. However, we do not categorize a paper as a discussion paper merely because the abstract mentions that it contains discussion or will consider implications, as they are typical elements of application papers and do not qualify the paper for the discussion category unless they represent its main research emphasis.

In terms of model architecture, this two-stage approach balanced performance with time and budget constraints and helped to ensure replicability. In the first stage, we randomly sampled 500 AI-related publications (including titles and abstracts) from each of the 19 scientific domains and used zero-shot prompting with GPT-4o to generate annotated data based on the concatenation of title and abstract to classify the AI mode into either Application or Discussion. The prompt used to conduct classification of AI mode is shown in Fig. 2. In the second stage, we fine-tuned a SciBERT model on the 9500 GPT-labeled examples.Footnote 3 This fine-tuned SciBERT model was then used to classify all publications contained in All AI Publications, with, 698,048 publications (88.8%) classified as Application, and 88,341 (11.2%) classified as Discussion.

Fig. 2
Fig. 2
Full size image

Prompts used to classify application and discussion AI publications. Note: Illustration of prompt used with GPT-4o to classify publications into either Application or Discussion. See text for explanation and details of subsequent processing

Empirical model

To explore the relationship of GenAI adoption on collaborations, we construct measures of GenAI, which is a dummy variable coded 1 if a publication is related to GenAI and 0 for Other-AI. The variable of team size (TS) for each publication is calculated as the number of authors listed on each publication. The variable of number of countries (NC) is calculated as the number of countries represented by the authors’ affiliations.

We estimate two models to explore the correlation between GenAI adoption and team size and number of countries of research correspondingly. Both measures are count variables and we therefore estimate Poisson models as follows:

$$\text{ln}\left({TS}_{i}^{\left(1\right)}\right)={\upbeta }_{1}GenA{I}_{i}^{\left(1\right)}+{\sum }_{t=1}^{T-1}{\upgamma }_{t}{\text{Year}}_{t}+{\sum }_{j=1}^{J-1}{\updelta }_{i}{\text{Subject}}_{i}+{\upepsilon }_{i}^{\left(1\right)}$$
(1)
$$\text{ln}\left({NC}_{i}^{\left(2\right)}\right)={\upbeta }_{2}GenA{I}_{i}^{\left(2\right)}+{\upbeta }_{3}{\text{ln}(TS}_{i}^{\left(2\right)})+{\sum }_{t=1}^{T-1}{\upgamma }_{t}{\text{Year}}_{t}+{\sum }_{j=1}^{J-1}{\updelta }_{i}{\text{Subject}}_{i}+{\upepsilon }_{i}^{\left(2\right)}$$
(2)

where GenAI indicates whether the publication is classified as generative AI. TS is the number of authors for each publication \(i\). NC is the number of countries for each publication \(i\). As usual, \(\upepsilon \_i\) is the error term while \(\gamma \_t\) and \(\delta \_i\) stand for year fixed effects and subject effects respectively to account for year-specific influences (ranging from 2017 to 2023) and disciplinary differences (with the subject score assigned by using the confidence score of the 19 root-level concepts in OpenAlex).

Results

Publication trends of GenAI across disciplines

Figure 3 illustrates the annual trend in overall GenAI and Other-AI scientific publications as well as the disciplinary-specific publication trends for All AI Publications and AI Application Publications. Plots of publication numbers are on a logarithmic scale. We observe that the yearly trend of GenAI and Other-AI related publications closely mirrors the trend of AI applications overall. However, the gap between All AI and AI Application Publications is more pronounced (i.e., there are relatively more discussion papers) in certain social science domains, such as political science, sociology, philosophy, business, and art. Yet, while relatively more prevalent in social science domains, AI and GenAI discussion papers are found across the disciplinary landscape of science.

Fig. 3
Fig. 3
Full size image

Source: Analysis of OpenAlex publication metadata. Number of records for All AI Publications: GenAI = 10,395; Other AI = 775,994. Number of records for AI Application Publications: GenAI = 6303; Other AI = 691,745. A logarithmic scale is applied to the Y-axis. The field of study for each publication is determined by selecting the top OpenAlex domain with the highest confidence score

Overall and disciplinary-specific publication trends, 2017–2023.

Additionally, while Other-AI publication output grew steadily from 2017 to 2023, GenAI applications in science experienced even faster growth, with a notable surge beginning in 2022. This exponential growth pattern reflects the recent surge of engagement with generative AI. GenAI applications have boomed in computer science, although Other-AI related research in computer science continues to be far greater (10×) in scale. A similar pattern is seen in the growth and scale relationships of Other-AI and GenAI in medicine. In fields such as art, sociology, and psychology, GenAI has captured a relatively greater share of AI publications compared with science and engineering domains. Possibly this reflects ease of entry, debate about its societal implications, and attention to implications for education and cognitive development (e.g., in psychology). Overall, the wide-ranging reference to GenAI in publications across diverse scientific domains highlights its cross-disciplinary nature and its potential to influence change in approaches, practices, and methods across multiple research fields.

Geographical distribution of GenAI

Figure 4 presents the publication output of the top 20 most productive countries in GenAI and Other-AI research from 2017 to 2023, for All AI Publications, AI Application Publications and AI Discussion Publications.Footnote 4 Each bar shows the percentage contribution of a given country or region to global GenAI and Other-AI publications. The United States and China are the two leading countries in AI-related research and active adopters of AI technologies, together accounting for a substantial share of global publications in both GenAI and Other-AI domains. The United States leads in producing GenAI-related research publications with 39.1% of global output, while China dominates in Other-AI-related research with 30.7% of worldwide publications. The contrast between GenAI and Other-AI publications in the United States and China highlights that high publication outputs in broader AI research, as exemplified by China’s leadership in Other-AI, does not necessarily translate to a leading role in GenAI research. In the emerging field of GenAI, the United States has decisively outpaced other nations in its publication output share through to 2023. This finding underscores the pivotal role of researchers in the United States in driving the rapid advancement of GenAI technologies.

Fig. 4
Fig. 4
Full size image

Source: Analysis of OpenAlex publication metadata. Number of records for All AI Publications: GenAI = 10,395; Other AI = 775,994. Number of records for AI Application Publications: GenAI = 6303; Other AI = 691,745. Number of records for AI Discussion Publications: GenAI = 4092; Other AI = 84,249. Leading countries are determined by All AI Publications

Leading countries for GenAI and Other-AI Publications, 2017–2023.

The next set of countries, after the US and China, by contributions to global AI publications include the UK, Germany, India, and Canada. Their shares of global Other-AI Publications range from 9.2% (for India) to 3.4% (for Canada). Several of these countries have relatively higher shares of global GenAI Publications, topping out at 7.8% for the UK. Moreover, when considering the ratio of world publication output shares in GenAI to those in Other-AI, Hong Kong and Singapore stand out, with ratios of 2.7 and 2.4, respectively. The US ratio, third by this measure, is 2.1, followed by the United Arab Emirates (1.7), Switzerland (1.7), Austria (1.4) and the UK (1.4). In other words, a set of economies (including smaller ones as well as the much larger US) have seen relatively higher shifts in their AI publication outputs towards GenAI. Mainland China (ratio: 0.6) and India (0.5) have exhibited relatively lower shifts, placing them towards the bottom of the rankings among all countries. Also of note is China’s contribution to the world’s GenAI discussion papers which, at 9.6% is much lower than for the US (34.3%) and just above the UK (8.4%). These results highlight not only the global landscape of AI research but also the evolving dynamics between GenAI and Other-AI fields. It remains to be seen how the observed patterns reflect broader trends towards specialization in AI research and how these patterns might evolve in the face of new technological opportunities in the AI landscape.

While most research publications in AI domains are produced by universities, industry has a significant role in AI research (Ahmed et al., 2023), which may also relate to observed geographical differences and is important when considering access, equity, and other ethical questions. Institution type variables are available in our dataset from OpenAlex, sourced from the Research Organization Registry (ror.org). The types include categories such as education, company, or government. Overall, about 20% of GenAI research publications are industry authored or co-authored, compared with 8% for Other-AI publications. The top industrial companies for applying AI technologies (including both GenAI and Other-AI) are in the US or China. This roster is led by Google (US) with 947 AI application publications, followed by Shanghai Electric (933), Tencent (832), Samsung (801), China Southern Power Grid (675), Alibaba Group (656), Microsoft (653), Huawei Technologies (633) and Intel (605). We report descriptive statistics for industrial GenAI and Other-AI application publications by three categories: (a) academic-industry collaborations; (b) authored solely by industrial researchers; and (c) all application publications involving industry. The latter comprises not only (a) and (b) but also industry application publications authored with non-academic partners (Fig. 5). For GenAI, there was a significant early growth in industrial contributions to AI application publications, peaking at 30.1% in 2020 for all industry authorships and 15.6% in 2019 for sole industry authorships compared with 8.6 and 1.8% respectively in the same years in industry contributions to Other-AI application publications. Industry contributions to GenAI application publications, although higher absolutely, were relatively lower as a share of all publications in the 2021–2023 period, reflecting a greater expansion of university outputs following early GenAI industrial leadership. Over the same time period, industry contributions to Other-AI application publications were relatively stable and lower.

Fig. 5
Fig. 5
Full size image

Source: Analysis of OpenAlex publication metadata. Number of records for AI Application Publications: GenAI = 6303; Other AI = 691,745. *Total of all industry authorships, including academic and non-academic co-authors

Industry involvement in AI Applications, 2017–2023.

Regression results

To investigate the correlation between GenAI use and collaboration, we estimate the two equations presented in the empirical model section (above).

We note that the skewness and kurtosis values of the variables reveal that Team Size (TS) is highly skewed (> 300) and exhibits excess kurtosis (> 200 K), indicating the presence of numerous extreme outliers. To address this, we remove the outliers by applying a 99th percentile threshold. In the All AI Publications dataset, a total of 7237 outliers were removed, including 139 GenAI and 7098 Other-AI publications. In the AI Application Publications dataset, 6661 outliers were removed, comprising 99 GenAI and 6562 Other-AI publications. In the AI Discussion Publications dataset, 748 outliers were removed, including 53 GenAI and 695 Other-AI publications.

Table 3 provides descriptive statistics for the key dependent variables across the three samples. In the All AI Publications sample, GenAI-related papers have an average of 4.04 authors and involving 1.33 countries, whereas Other-AI-related papers have 4.15 authors and involve 1.30 countries. For AI Application Publications, the differences in team size and international collaboration are more pronounced: GenAI papers have an average of 4.61 authors and involve 1.33 countries, while Other-AI papers have an average of 4.25 authors and involve 1.30 countries.

Table 3 Descriptive statistics
Full size table

The baseline regression results on the three samples (All AI, AI Application and AI Discussion) are presented in Table 4. The dependent variable TS is shown in columns (1), (3), and (5), with the Number of Countries (NC) shown in column (2), (4), and (6). We estimate a Poisson model with year and subject fixed effects, reporting marginal effects with robust standard errors. The results are robust to OLS estimates using logged TS/NC values and negative binomial estimates. Figure 6 shows the estimated margins of TS and NC for GenAI and Other-AI application publications.

Table 4 Regression results of GenAI diffusion on team size (TS) and international collaboration by number of countries (NC)
Full size table
Fig. 6
Fig. 6
Full size image

TS and NC estimation by Poisson regression controlling for years and subjects. Note: TS = Team Size; NC = Number of Countries. The error bars represent the 95% confidence intervals. Two-tailed test: *p<0.05, ***p < 0.001

The results show that for AI Application Publications, the marginal effect of GenAI is positive and significant for TS (column 3), suggesting a positive correlation between GenAI applications and team size, compared to Other-AI applications. Figure 6 graphically shows that GenAI application publications tend to have bigger team sizes compared to Other-AI application publications (4.57 vs. 4.25, p < .001). One possible interpretation is that application-oriented research with GenAI, possibly due to its newness, requires additional technical and domain-specific expertise, leading to larger teams compared to more established AI techniques. Yet, caution must be used when interpreting these results, as the observed effect is small, and teams remain relatively large in both cases (mean of > 4 authors).

By contrast, in column (4) Table 4, we can see that GenAI is positively related to the number of countries for AI Application Publications, indicating that GenAI applications involve more international collaboration compared to Other-AI research. This is presented graphically in Fig. 6 (1.31 vs. 1.30, p < .001). This suggests that GenAI applications are associated with a more globally connected network, possibly due to the collaborative nature of cutting-edge GenAI applications and the shared international interest in GenAI technologies. Again, the differences are small, and the majority of publications do not involve international collaboration for both GenAI and Other-AI publications (median = 1 country). For Gen AI Discussion Publications compared with Other-AI Discussion Publications, team size is slightly smaller, but this difference is not significant. GenAI Discussion Publications are also associated with slightly more author countries than Other-AI Discussion Publications. Overall, while modest, the divergence between team sizes and international collaboration warrants further investigation, particularly in exploring whether it reflects differing research motivations, the accessibility of GenAI tools, or broader geopolitical trends in AI research.

Discussion and implications

Our exploratory analysis of all AI publications published between 2017 and 2023, reveals that reference to GenAI in scientific research has expanded well beyond its origins in computer science. While early developments in GenAI were predominantly concentrated within the computer science field, we now observe the broader diffusion of GenAI concepts, models, and approaches across a diverse range of scientific disciplines. This includes fields such as medicine, chemistry, geography, and even sociology, where researchers are increasingly leveraging GenAI for tasks such as data generation, predictive modeling, and hypothesis testing. We confirm this by developing and deploying a classification model that distinguishes between publications that apply AI methods and those that focus on discussing AI without implementing them. The cross-disciplinary growth of GenAI application publications, suggests that GenAI is emerging as a general purpose technology (Eloundou et al., 2024) and is becoming a tool used not just within its foundational domain, but as an innovation for enhancing research methodologies, accelerating discovery, and addressing complex scientific challenges in a variety of fields.

We find that the US has moved more rapidly into using GenAI in publications than China (through to 2023). The US research community is both contributing to and benefiting from US leadership in GenAI innovation and implementation. This is supported by a dynamic tech ecosystem, high R&D investment, substantial funding from both major tech firms and venture capital, and policies prioritizing advanced AI R&D (Fattorini et al., 2024; NSTC, 2023). China’s high volume of Other-AI publications highlights its broad R&D drive, strong investment in mainstream AI, national policy support, and a co-evolved public–private AI ecosystem (Lundvall & Rikap, 2022; State Council, 2024). The Chinese research community was slower in adopting Gen-AI through 2023. However, looking forward we can see that China is now intensifying GenAI efforts, especially in the private sector, and developing its own models such as Deepseek (2025). Notably, we find that China’s share of GenAI application papers is substantially higher than its share of GenAI discussion papers, compared to other leading countries. This suggests a research culture that is more focused on implementation and perhaps more optimistic about the benefits of GenAI, with less visible engagement in critical reflection about its broader implications. This pattern highlights the importance of examining not only whether GenAI is being referenced, but also in what capacity—tool or object of study—as this distinction may reflect national differences in institutional incentives, policy orientations, and epistemic cultures.

Concerns remain for China, as well as for the US and elsewhere, about the openness and governance of GenAI (Chang, 2024; Taeihagh, 2025; Triolo & Schaefer, 2024). Researchers in numerous other economies are also engaging with GenAI in their publications, with relatively high levels of deployment seen in several small and medium-sized advanced economies. Industrial involvement in the growth of GenAI research publications has been significant, most notably in earlier years of our study period (up to 2020).

Our regression results suggest that for the categories of papers that we examined, GenAI Application Publications tend to be authored by larger teams and involve slightly more author countries compared to other types of AI application publications. Application-oriented research with GenAI, possibly due to its novelty, may require additional technical and domain-specific expertise, necessitating larger teams, with increased global cooperation reflecting widespread interest in the approach. Interestingly, this finding runs counter to common narratives about AI and automation unidirectionally leading to labor substitution or downsizing. Instead, our results are more nuanced, suggest that GenAI—at least in its current state—may be complementing rather than replacing scientific labor. Indeed, its application may introduce new complexities and capabilities that require coordination and additional expertise across broader interdisciplinary teams compared to more established AI technologies. It is also possible that those already working in larger, more distributed teams are better placed to utilize new technologies and act as lead users. In other words, the application of GenAI in science may amplify rather than diminish the need for collaborative, collective research efforts.

GenAI Discussion Publications appear not to need larger technical teams, as there is a hint that they might be written by slightly smaller teams than OtherAI Discussion Publications (although this is not statistically significant when controlling for other factors). Still, GenAI Discussion Publications involve authors from a greater number of countries compared to other AI Discussion Publications, conceivably reflecting widespread international interest about the implications of GenAI.

Our findings contrast with the study by Thu et al. (2022), which examines team sizes in machine learning (ML)-related and ML-unrelated projects. Their research indicates that ML-related projects tend to involve slightly fewer authors compared to ML-unrelated ones, with no significant difference in the number of countries represented. They suggest this might be attributed to the reduced need for physical work in ML-related projects, thereby requiring fewer team members. As GenAI further evolves and diffuses, researchers may benefit from more efficient workflows, automated processes, and enhanced model capabilities. These may help streamline tasks such as data generation, model training, and experimentation, although it is also possible that productivity effects could be dampened by research task amplification, additionality and diversification (Ribeiro et al., 2023). Additionally, despite the trend toward bigger team sizes, we find that GenAI research is slightly more international than for Other-AI. This is noteworthy in that it suggests that the level of international cooperation in GenAI research remains at least on par with, if not somewhat exceeding, the levels found in Other-AI fields notwithstanding the extensive array of geopolitical tensions that have arisen coincident with GenAI’s rapid growth in recent years.

We acknowledge that our study has limitations, including those related to definitions and data, although we have sought to be careful in implementing our search approach and cleaning and classifying the data. We also recognize that there are a variety of ways that scientists are using GenAI in their research publications as well as in research practices not necessarily captured by publication metadata. Our study, while exploratory, does indicate that GenAI is being increasingly adopted by scientists in their research across the multiple fields of science. This extends to this paper, where we demonstrate the capacity of GenAI as a classification tool in bibliometric research. Our two-stage classification model enables us to distinguish whether researchers are implementing GenAI tools or simply reflecting on their implications, offering a finer-grained view of how this transformative technology is shaping the research landscape. The net productivity effects of GenAI’s increased diffusion in science and how it will influence team sizes and international collaborations—as well as its implications for creativity, reproducibility, ethics, and other concerns—clearly all represent important topics for future research that can build upon these early findings.