The growing concentration of national influence in global science Authors: Charles J Gomez Published: 2026 ABSTRACT AbstractScientific influence, or the capacity of ideas and concepts to shape future research, is crucial to developing and disseminating knowledge and sustained innovation. Here, using nearly 240 million scientific works published between 1990 and 2023 from OpenAlex to construct international networks of influence, discursive influence, which represents what global scientific communities consider important and worthy of investigation, is shown to be disproportionately and increasingly concentrated in a small group of resource-wealthy countries, including the United States, Canada, Western Europe and East Asia, in comparison to attributional influence. This concentration raises issues of equity and innovation in global scientific discourse, perhaps narrowing research perspectives, exacerbating biases and creating echo chambers that are associated with stifling innovation and marginalizing contributions from countries that are peripheral to global scientific discourse. The findings underscore the need to promote diverse and inclusive global research enterprises.Subject terms: Scientific community; Science, technology and society; Sociology Main Influence in science occurs when ideas and concepts shape the direction, focus and evolution of future research and thinking. Such influence encompasses the ways in which scientific knowledge is created, disseminated, validated and built upon in scientific communities. It manifests in the publication of groundbreaking research that reflects the development of new theories or methodologies and even brings forth new paradigms. However, growing concern suggests that new scientific influence is disproportionately and increasingly emerging from a few resource-rich countries1,2, and thus researchers from only a handful of places are disproportionately influencing global scientific discourse3–6. This might merely reflect science as a successful meritocracy—researchers worldwide sifting through new ideas and concepts, rewarding only the most promising and best ideas. Such a concentration might reflect a reward system of attention and recognition across fields. However, science is structurally plagued by the Matthew Effect7, wherein early gains of attention, resources and recognition yield disproportionate rewards, decoupled from merit. This is concerning for two reasons. First, sustained scientific innovation might be in jeopardy. Research into collective problem-solving warns that even if ideas and collaborators are of greater quality, monopolized attention on only select ideas or collaborators might be associated with inhibiting the diverse perspectives and ideas needed for serendipity and novel combinations of ideas1,8,9, which might impact sustained innovation across fields. Second, there is concern regarding ensuring equity and inclusion of peripheral perspectives in science. Scientists from traditionally underrepresented, if not excluded, countries are increasingly participating in global scientific conversations. Such scientists are more likely to internationally coauthor research and publish in internationally recognized journals10–14, and yet the reach of their ideas might be far more limited. Influence is embodied and measured empirically using scientific output, or metadata that associate with scientific publications, the medium of discourse and the evolution of the pool of epistemic knowledge in science. Influence in science is also often conceptualized and understood as a singular measure, and yet little attention is paid to how influence manifests in these scientific outputs, particularly internationally. Influence is deployed predominately internationally in bibliometric analyses in two ways: (1) influence as discourse and (2) influence as attribution. Discursive influence15 represents how ideas and concepts that are embodied by scientific publication terms or phrases diffuse across the scientific community’s worldwide discourse by being incorporated in papers. Text data are the ideal medium to embody scientific knowledge’s abstract and complex nature, where it can be extracted and measured at scale1,15–20, making discursive influence a highly useful and attractive way to gauge influence. When text is extracted from abstracts or titles of papers, discursive influence distinctly reflects the focus of influencing papers and influenced papers. By examining discursive influence, ideas and concepts that are gaining attention and shaping the direction of scientific enquiry can be identified. It provides insights into the intellectual currents in the scientific community, and it highlights trends and shifts in research priorities. In contrast, attributional influence represents the recognition of work in international scientific communities through citations that reflect the networks and paths of established knowledge and those who have established it. Citations as attributional influence reflect intellectual acknowledgement of an idea’s impact on the field or its author who developed the idea21–25. Attributional influence reveals paths of established knowledge, demonstrating how ideas are built on and which researchers or works are foundational. Attributional influence helps identify and understand the intellectual lineage of scientific ideas, as well as identify contributors to the advancement of knowledge. Discursive and attributional influences are complementary; they demonstrate how scientific ideas not only proliferate, but gain acknowledgement, highlighting the intellectual lineage of ideas and their diffusion across global scientific discourse. Attributional influence, through citations, reflects the breadth of ideas incorporated from various sources for diverse purposes, and discursive influence represents focused research priorities. Despite its ability to shape academic discourse, the latter receives less attention, raising concerns about its influence on research trends and priorities. Parsing and understanding influence as either discursive or attributional reveals divergences between the two at the global level, with consequences not only for the growing concentration of international influence and inequality, but for sustained innovation of the scientific enterprise itself. Using nearly 240 million papers from a public dataset of scientific publication metadata called OpenAlex26, thousands of potential scientific concepts, ideas and phrases, and the papers that mention them in either their abstracts or titles, from 1990 to 2023, are identified. All works that cite these papers (that is, attributional influence) and those that not only cite them but also mention the same terms in their abstracts or titles (that is, discursive influence) are identified. International networks of attributional and discursive influence are constructed, finding that they correlate greatly and supporting that both are complementary. A growing body of research suggests that scientific influence is increasingly concentrated in a small number of ‘core’ countries1,6,27. These core countries—including the United States, Canada, Western European nations and parts of East Asia, such as China—are typically resource-rich nations that house the world’s top universities and produce the highest volume of scientific research. China’s inclusion as a core country reflects its considerable rise over the past few decades to become a major contributor to science and technology28–30. In contrast, ‘periphery’ countries are those with fewer resources for scientific infrastructure and are often underrepresented in global scientific discourse. While researchers from the periphery are increasingly participating in international science, their work may be overlooked. This dynamic creates a stratified global scientific community where a country’s position in the core or periphery can impact the visibility and influence of its research. Where influence originates reveals a growing global imbalance. Specific paths of ‘relative influence’ between attributional and discursive influence reveal not only a growing disconnect between the two but also international concentration within them. The results show that attributional and discursive influence increasingly come from core countries. In these locations are the best universities that publish the most research, in comparison to those that lie on the periphery. These countries are increasingly and disproportionally concentrating influence among themselves, deciding what is important and worthy of research in the global scientific community, rather than bridging the growing divide between core and periphery countries. With growing concentration of discursive ‘relative’ to attributional influence, core countries might disproportionately become trendsetters in scientific research, deciding the direction of future enquiries and innovations. Core countries might also increasingly shape the global research agenda by promoting specific terms, concepts and methodologies, leading to a focus on issues more relevant to them and potentially neglecting problems pertinent to less influential areas. This might result in a homogenized scientific landscape, in which diverse research interests and perspectives are underrepresented. Such growing concentration of influence has implications not only for equity and inclusion but also for sustained novelty and breakthroughs in science31. Discovery is predicated on a diversity of ideas and perspectives and serendipity. Growing global conformity in scientific research from a few countries might constrain sustained innovation and threaten the scientific enterprise. In contrast, with growing and greater concentration of attributional ‘relative’ to discursive influence, periphery countries might see their research cited widely among themselves, not necessarily shaping leading scientific discourse. This might lead to local research priorities and contexts being overlooked in favour of topics deemed important and not integrated into current and emerging scientific conversations, limiting its immediate impact on the direction of new research. Results Constructing discursive and attributional influence Much attention is paid to the concentration of attributional influence and its worrying trends2, but the concentration of discursive influence is overlooked, even though it suggests a more concerning trend. Attributional influence is defined using citations in a paper, which serve various purposes. They acknowledge extant work, support a paper’s arguments and provide context. This type of influence is characterized by wide distribution across numerous sources and reasons for inclusion32,33. Citations are scattered throughout a document, integrated for various reasons, and they cater to disparate objectives, making attributional influence dispersed and multifaceted. In contrast, discursive influence pertains to using terms and concepts in the abstracts or titles of academic papers that reflect their foci. Although they might appear in the body of a paper, terms in an abstract or title signify a paper’s research focus, encapsulating the essence of a study’s central theme. Unlike attributional influence, which amalgamates ideas from existing works, discursive influence assesses the present study’s main concerns and objectives. This focal specificity makes discursive influence more troubling when overlooked, since it shapes the research agenda and the direction of scholarly discourse directly. When countries are more discursively influential ‘relative’ to their attributional influence, it suggests that their research agendas, priorities and frameworks are shaping global scientific discourses disproportionally. This raises several concerns. First, it results in a narrowing of perspectives. If a few countries dominate discursive influence in academic literature, it can lead to a homogenization of research priorities. Such dominance means that unique challenges, perspectives and innovations from less discursively influential countries are underrepresented or overlooked. It also limits the diversity of ideas and approaches that are crucial to addressing complex global issues. Second, it exacerbates biases in research priorities. Discursive influence shapes what is considered important and worthy of investigation in academic communities. If some countries exert greater discursive influence, there is a risk that research agendas become biased towards their interests, concerns and viewpoints. Such bias affects funding allocation, publication opportunities and the overall direction of scientific enquiry, potentially ignoring issues that are critical to other regions of the world. A concentration of discursive influence also leads to the creation of echo chambers, in which ideas, methodologies and theories are reinforced and alternative viewpoints are marginalized. Such reinforcement stifles innovation and critical thinking, leading to less dynamic and responsive academic environments. Disproportionate discursive influence from a few countries might skew policy recommendations and interventions towards those that favour influential countries’ contexts, ignoring the nuances of local contexts in less influential regions. Researchers in areas peripheral to mainstream scientific communities might find it increasingly difficult to break into established discursive frameworks if they are dominated by a few. This creates barriers to entry for new ideas and researchers, especially those from underrepresented regions, making it more difficult for them to gain recognition and contribute to global scientific discourse. Attributional influence represents the breadth of sources and ideas recognized through citations, and a higher ‘relative’ discursive influence signifies a concentration of power in shaping what is studied, discussed and valued in global science. Such a ‘relative’ imbalance poses challenges to the global research enterprise’s diversity, equity and inclusiveness and its ability to address global challenges with nuanced, varied perspectives. To this end, a majority-vote ensemble of several popular key phrase extractors written in Python 3.8 that isolate phrases with at least three characters, ranging from unigrams (singular terms) to five-grams (five consecutive terms) is used, from the titles and abstracts of 239,036,158 million papers published between 1990 and 2023 from OpenAlex. All papers that incorporate the term in their abstracts or titles are identified, and these papers are called ‘term papers’. Only terms whose first appearance in the data is at least in the year 1990 and appears in at least 10 ‘term papers’ in OpenAlex are considered. ‘Term papers’ must have been cited to be included in the analysis. In the Discussion and ‘Limitations and further research’ sections, the advantages of this approach over considering only the first paper in the data in which a term appears are discussed, reflecting the introduction of a novel idea or concept, since this approach is used in other studies of influence34. English is largely the common language of scientific publications, evidenced over the past 20 years in many scientific fields35–37, so papers published in English are considered when identifying and extracting terms. This choice is discussed in detail in the Methods, Discussion and ‘Limitations and further research’ sections. For each ‘term paper’ i, all future works that cited it within a 10-year window after publication are identified. A limited timeframe is used because the influence of a specific paper is most direct within a certain period after its publication. Over longer periods, researchers may use terms from older papers without citing them, as such terms have become common knowledge in the field. A 10-year window was chosen as a conservative measure to capture the peak of a paper’s influence while minimizing the noise from concepts that have become broadly integrated into the scientific lexicon. Citing papers that use the same term in their abstracts or titles as the cited ‘term paper’ i are called ‘discursive papers’. Works that merely cite ‘term paper’ i but do not have the term in their abstracts or titles are called ‘attributional papers’, reflecting how attributional influence through citations acknowledge the ideas in cited papers and employ them in the various ways outlined. In Supplementary Note 1, how ‘attributional’ and ‘discursive papers’ are identified to construct attributional and discursive influence networks, respectively, is further outlined. This distinction is important. A paper merely having the same term in its abstract or title is not a clear indication of discursive influence, but when a paper cites the ‘term paper’ that has the same term in its abstract or title (within the 10-year time window), it is a strong indication of discursive influence. ‘Discursive papers’ and ‘attributional papers’ are mutually exclusive; a citing paper that shares the same term as the ‘term paper’ is counted as a ‘discursive paper’, not an ‘attributional’ one. ‘Discursive papers’ themselves are also ‘term papers’ that associate with their own set of ‘attributional papers’ citing it, and their own set of ‘discursive papers’ citing it and with the shared term in its abstract or title. This process was repeated for all subsequent ‘term papers’ for all terms, resulting in 10,046 terms that appear in over half a million (567,258) ‘term papers’ that originated from 213 countries. Figure 1 reports descriptive statistics and trends of the ‘term papers’ used to construct international networks of influence. Figure 1a plots the per cent of ‘term papers’ on the y axis from the top 20 countries that authored the most ‘term papers’ on the x axis, ranked left to right from highest to lowest percentage. These countries represent over 80% of all ‘term papers’ used during analysis that met various criteria (for example, a term that appeared in at least 10 papers, a term that first appeared after 1990 and so on.). Figure 1b plots the distribution of the number of ‘term papers’ (that is, the papers with a particular term in its abstract or title) on the x axis, with the number of ‘attributional papers’ per term (that is, the number of citations these ‘term papers’ received within 10 years of their publication) on the y axis. Both axes use a logarithmic scale (log10) to represent the range of values. Terms are represented as data points but visualized using a hexagonal heat map, in which lighter colours indicate higher frequencies, showing where clusters of terms, as data points in the plot, are densest (Pearson correlation coefficient of r = 0.78, P < 0.001, 95% CI 0.77–0.78). Histograms independently display the distribution of terms along both axes. Multiple examples and details of terms are discussed in Supplementary Note 2. The average number of ‘term papers’ per term is 67, meaning that the average number of papers with the same term is 67. The average number of ‘attributional papers’ per term (the number of citing ‘attributional papers’ that ‘term papers’ received within the 10-year window for a given term, averaged across terms) is 1,933 papers. Thus, for each term, the number of ‘attributional papers’ citing its ‘term papers’ is 1,933. The distribution of term papers. a, Per cent of term papers (y axis) from the top 20 countries that authored the most ‘term papers’ (x axis). b, Number of ‘term papers’ per term (x axis) and the number of attributional papers per term, or the number of citations of term papers within 10 years of it being published (y axis). Both the x and y axes are plotted on a log10 scale. Data points are terms visualized as a hexagonal heat map, where lighter colours reflect higher frequency. Histograms are plotted on the margins. Statistical test (b): Pearson correlation (two-tailed), r = 0.78, P < 0.001, 95% CI 0.77–0.78, n = 10,046 terms. Quadratic assignment procedure models of international networks of influence The analysis shows that international networks of attributional and discursive influence are highly correlated. A linear regression model for networks, called quadratic assignment procedure (QAP), was used to calculate this correlation. Yearly international networks of attributional and discursive influence from 1990 through 2013 were constructed. This is shown in greater detail in Supplementary Note 3. The analysis was stopped at 2013 since a 10-year window was used, and the version of the data that were used from OpenAlex ends in 2023. The international networks for attributional influence were constructed using ‘term papers’ and ‘attributional papers’, or papers that cite the ‘term paper’ but do not have the same term in the abstract or title. They were constructed as the number of ‘attributional papers’ that cite a ‘term paper’ published within the 10-year, post-publication time window. The countries that are associated with the ‘term paper’ and those that are associated with the ‘attributional paper’ are noted. Each carries a weight of ‘1’. In cases of internationally co-authored papers, the weighted edge of ‘1’ is split among the countries proportionately. These directed weighted international edges are combined and summed across all ‘term papers’ and their associated ‘attributional papers’ each year, creating the international network of attributional influence. In Supplementary Note 4, the edge weights are normalized for both attributional influence and discursive influence by terms, where the edge weights are percentages per term. Then, the percentage edge weights are summed by country dyads in each year. The results are the same. Figure 2 plots β coefficients from 24 QAP models (that is, one model per year), starting in 1990 and continuing to 2013. The dependent variable is the international network of discursive influence for year t, and the independent variable is its corresponding international network of attributional influence for year t. To render the two networks comparable across time and within a year, the edge weights are converted to z-scores (that is, mean-centre and network edge weights normalized by standard deviations), which converts the QAP intercept coefficient to null, allowing comparisons across models. Each model produced a β coefficient that reflects how a one-standard-deviation increase in attributional influence corresponds to a change in discursive influence, measured in standard deviations. The QAP model also produces standard errors, shown by the shaded area, that are associated with each β coefficient, which were statistically significant at a two-tailed 0.05 significance level. QAP model of attributional influence and discursive influence over time. β coefficients from the QAP regression models of international networks of discursive influence regressed onto international networks of attributional influence plotted by year from 1990 to 2013. On the y axis are the β coefficients (interpreted as z-scores), plotted as data points, with the shaded region as the standard errors for each year on the x axis. Data points represent yearly QAP regression β coefficients (mean of bootstrapped distribution); shaded region shows ±1 s.e. The red dashed line indicates parity (β = 1). Statistical test: QAP regression with 10,000 permutations, two-tailed; all yearly coefficients significant at P < 0.001. Correlation between β and year: r = 0.75, P < 0.001, 95% CI 0.49–0.88. Mean β across years = 0.98 (95% CI 0.97–0.99). The models demonstrate that discursive and attributional influences associate highly with one another as international networks; countries that influence attributionally also influence discursively in equal measure. A value of 1 reflects parity; one standard deviation in attributional influence corresponds to one standard deviation in discursive influence, shown in Fig. 2 as the horizontal dashed red line that intercepts the y axis at 1. With a Pearson correlation coefficient of r = 0.75 (P < 0.001, 95% CI 0.49–0.88), the average value of these β coefficients across time is 0.98 (95% CI 0.97–0.99), meaning that a one-standard-deviation increase in the weighted edges of an attributional influence international network associates, on average, with a one-standard-deviation increase in the weighted edges of its corresponding discursive influence international network. This is noteworthy because international discursive influence networks were constructed from a mutually exclusive set of papers (that is, ‘discursive papers’) from international attributional influence networks (that is, ‘attributional papers’). This high correlation suggests an apparent parity at the aggregate global level; a country’s attributional influence seems to move in lockstep with its discursive influence. However, this macro-level convergence presents a puzzle, as it masks growing divergences when influence is examined between specific groups of countries, a dynamic explored in the following sections. ‘Where’ influence originates reveals a growing global imbalance. To explain this contradiction, the ‘type’ of country is used in both international networks of attributional and discursive influence. Figure 3 plots the centralities of countries in the international networks of discursive and attributional influence. For attributional influence, a country’s centrality score in the network is its degree centrality, or the number of works that cite a paper with a term within the 10-year time window after it was published and scaled yearly as a z-score. For discursive influence, a country’s centrality score in the network is its degree, or the number of works that cite a paper with a term within the 10-year time window after it was published that has the same term in its abstract or title, also scaled yearly as a z-score. Centralities of countries and their concentration in international networks of attributional influence and discursive influence plotted by year from 1990 to 2013. a,b, Gini coefficient (y axis) among country centralities in both international networks of attributional influence (a) and discursive influence (b) over time (x axis). c, Average country centralities (y axis, as z-scores) for both core and periphery countries in both networks over time (x axis). Statistical test: Pearson correlation (two-tailed) with n = 24 yearly observations. In a: r = 0.88, P < 0.001, 95% CI 0.74–0.95. In b: r = 0.86, P < 0.001, 95% CI 0.70–0.94. In c, data points represent yearly mean centrality values (mean of country centralities for each group); shaded regions show ±1 s.e.m. For core countries: attributional r = 0.98, P < 0.001, 95% CI 0.95–0.99; discursive r = 0.99, P < 0.001, 95% CI 0.95–0.99. For periphery countries: attributional r = 0.87, P < 0.001, 95% CI 0.74–0.95; discursive r = 0.86, P < 0.001, 95% CI 0.72–0.94. The Gini coefficient38, calculated across all centralities for each year, is a standard measure of inequality. A value of 1 reflects full inequality and 0 reflects perfect equality in an economy. The Gini coefficient is used to reflect whether influence is increasingly and disproportionately concentrated among a handful of countries. Since the coefficients require all values to be positive, the minimum centrality value is added to all centrality values for a year so that no value is negative. In Fig. 3a,b, Gini coefficients applied to the centralities for attributional and discursive influence are rising over time. Each data point is a Gini coefficient of the distribution of centralities. Figure 3a shows the coefficients applied to attributional influence network centralities, and Fig. 3b shows discursive influence network centralities. Both show that the concentration of influence increasingly originates from some countries in both networks over time (attributional influence with a Pearson correlation coefficient of r = 0.88, P < 0.001, 95% CI 0.74–0.95, and discursive influence with a Pearson correlation coefficient of r = 0.86, P < 0.001, 95% CI 0.70–0.94). Figure 3c shows the type of country in which concentration is most prominent. It plots the average centrality for core and periphery countries for both attributional and discursive influence networks over time. Each data point is the average centrality for core or periphery countries each year. Core countries appear in blue, and periphery countries are in red. The left pane plots attributional influence and the right pane plots discursive influence. Core countries are increasingly central in both networks at a much higher rate than periphery countries (for core countries: attributional influence with a Pearson correlation coefficient of r = 0.98, P < 0.001, 95% CI 0.95–0.99; and discursive influence with a Pearson correlation coefficient of r = 0.98, P < 0.001, 95% CI 0.95–0.99; for periphery countries: attributional influence with a Pearson correlation coefficient of r = 0.88, P < 0.001, 95% CI 0.74–0.95; and discursive influence with a Pearson correlation coefficient of r = 0.87, P < 0.001, 95% CI 0.72–0.94). Both international networks of influence are increasingly unequal and concentrated in a few core countries. International networks of relative influence To resolve the contradiction between the high global correlation of influence types and the observed concentration of influence in core countries, it is necessary to isolate their meaningful divergences. To show this, ‘relative influence’ is defined as the difference between every dyad (that is, country-to-country) in the international networks of attributional influence and its corresponding dyad in the international network of discursive influence. The work on citational lensing1, which identifies growing deviations between what is being researched between countries and how infrequently they are being cited, is used. For year t, the dyad in the discursive network (Discursivei→jt), where country i is the national affiliation of the paper with the same term in the abstract or title as the citing paper affiliated with country j, is matched with the corresponding dyad in the attributional network (Attributionali→jt), where papers affiliated with country i receive citations from papers affiliated with country j. The weighted edge values of these dyads are subtracted across all dyads in these networks for year t to construct a ‘relative influence’ network, defined as: A positive value of Relativei→jt, where Discursivei→jt>Attributionali→jt, indicates that country i exerts more discursive influence on country j ‘relative’ to its attributional influence on country j. Conversely, a negative value of Relativei→jt, where Discursivei→jt