Structure of scientific knowledge flows to intergovernmental organizations

Significance

What scientific findings are incorporated into intergovernmental organization (IGO) documents and why? This study revealed that a small group of scientists accounts for a substantial portion of the research that is used by international organizations in their efforts to connect science with policy, a level of concentration comparable to that observed in leading journals. These scholars maintain dense international collaboration networks, experience rapid uptake in IGO documents, and are cited by multiple IGOs. A comparison between mature (e.g., climate modeling) and emerging (e.g., data science & AI) fields revealed that mature fields showed a greater concentration. This concentration persists over time despite IGOs’ efforts to enhance legitimacy by broadening participation in evidence synthesis.

Abstract

Intergovernmental organizations (IGOs) attempt to shape global policy through scientific guidelines and assessments. While they rely on external scientists to bridge research and IGO advisory processes, the structural pathways connecting science to IGO documents remain unexamined. By linking 230,737 scientific papers referenced in IGO documents (2015–2023) to their authors and coauthorship networks across 23 research fields, we identified a small cohort of “Highly IGO-Cited Scientists” (HIC-Sci)—typically comprising 0.7% to 4.4% of authors whose work accounts for 30% of IGO-cited papers. This structural concentration is associated with relational and cognitive patterns: dense transnational collaboration networks, overlapping memberships on advisory bodies such as the Intergovernmental Panel on Climate Change, rapid uptake in IGO documents, and standardized policy-oriented vocabularies. Geographically, HIC-Sci networks follow a core–periphery structure centered on Western Europe. Established fields, such as climate modeling, show stronger concentration, whereas emerging domains such as data science & AI show more distributed citation patterns. Major IGOs frequently cocite the same HIC-Sci papers, compounding this concentration through synchronized diffusion across IGOs. This concentration persists despite IGOs’ efforts to broaden participation and diversify their evidence base. While IGOs have developed criteria for selecting knowledge in advance, our framework provides a basis for subsequent assessment of how IGOs’ efforts to influence policy rely on a concentrated set of HIC-Sci.

Intergovernmental organizations (IGOs) are uniquely positioned at the interface of scientific research and policy-making to address global issues. Guidelines and assessment reports issued by IGOs, such as the Intergovernmental Panel on Climate Change (IPCC), the World Health Organization (WHO), and the World Bank, are widely referenced as the bases of both national policies and international agreements (1). However, IGOs do not directly produce scientific knowledge; instead, they rely on external scientists to translate research into policy-relevant terms. Cash et al. (2) argue that the effectiveness of science advice is dependent on an appropriate balance among salience (relevance to decision-makers’ needs), credibility (scientific adequacy of evidence and arguments), and legitimacy (perceived fairness and impartiality of knowledge production processes). Mitchell et al. (3) show that salience, credibility, and legitimacy interact in ways that often create trade-offs in global assessment processes. Empirical studies suggest, for example, that reliance on narrowly defined expert communities may enhance credibility while undermining perceived legitimacy (4, 5). This raises a critical question: Whose research is taken up by IGOs, and how concentrated is this uptake?

Citations have historically concentrated around a few scientists in academic research, a pattern long seen as a marker of trust (6). Merton’s account of cumulative advantage suggests that the concentration of recognition can facilitate deeper specialization and cumulative expertise (7–9). Similar dynamics have been theorized and termed the “accumulation of academic capital” by Bourdieu (10) and have been observed in preferential attachment within academic collaboration networks (11). Similarly, such concentration can support IGO knowledge selection. The Coupled Model Intercomparison Project Phase 6 (CMIP6) modeling community’s outputs, for example, now serve as shared reference points for climate policy (12), and reliance on established expert groups during the COVID-19 pandemic facilitated the rapid development of WHO guidelines (13).

However, such concentration can reinforce particular knowledge pathways through path-dependent dynamics (14, 15), shaping how scientific expertise is selected by IGOs to convey to policy makers (16). For example, the European hormone regulation (17), the IPCC glacier assessment episode (18), and the early WHO pandemic response (19) highlight how particular structural characteristics of expert networks can negatively affect policy-relevant knowledge selection. In response to such concerns, IGOs have adopted procedural reforms such as IPCC governance restructuring, WHO conflict-of-interest protocols (20), and Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services (IPBES) techniques to integrate different knowledge systems (21).

Despite these efforts, the nature of knowledge concentration itself has received little systematic examination. Prior work has examined the specific citations of individual papers in IGO documents (22, 23) and geographic biases at the country level (24), but because credibility and authority are often ascribed to individuals rather than to publications alone, author-level concentration warrants a dedicated analysis. It has been observed that author-level concentration has increased over time in scholarly communication (25), and authorship patterns in IPCC documents reflect the continued centrality of particular countries, institutions, and experts (26). However, such analyses remain largely descriptive and insufficiently comprehensive to clarify how influence on IGO document selection, which partly reflects credibility attributed to specific scientists, is constructed and retained through network structures, institutional roles, and cognitive alignment.

We examined the concentration of citations in IGO documents among specific scientists across three dimensions: structural, relational, and cognitive. Specifically, we tested three core hypotheses. First, at the structural level, citations in IGO documents concentrate on a relatively small number of high-profile scientists, consistent with preferential attachment dynamics (11). Second, at the relational level, these scientists, who occupy positions in both research networks and IGO advisory bodies, form dense collaboration clusters (27). Third, at the cognitive level, these scientists, who are aligned with IGO priorities, accumulate citations in IGO documents more rapidly, consistent with coproduction (28) and science advice as policy practice (29). By identifying these dimensions across research domains, we uncover systematic patterns concerning how scientific knowledge is incorporated into IGO documents.

We anticipated that these mechanisms will vary in intensity based on a field’s maturity. In mature domains, a convergence of scientific bases is likely to deepen reliance on particular expert communities, while in emerging fields, a more fluid knowledge base allows research from a wider range of scientists to be cited in IGO documents. We also examined whether cross-national and regional disparities in citations in IGO documents (24) are partly driven by a focus on specific authors.

Using a linked Overton–Scopus dataset of over 230,000 IGO-cited papers (2015–2023), we showed that citation in IGO documents is concentrated among a small group of scientists. Although citation in IGO documents does not directly measure policy uptake, it provides a tractable proxy for identifying scientists whose work IGOs select when synthesizing evidence. Our analysis revealed mechanisms sustaining this concentration: dense coauthorship clusters overlapping with advisory bodies (a case study in climate modeling), rapid institutional uptake, and preferential use of policy-aligned vocabularies. This concentration was observed across domains, but the intensity varied. Our results provide an empirical foundation for understanding the balance between the functional benefits of focused expertise and ongoing salience–credibility–legitimacy tensions in global science advice.

Factors Influencing Uptake of Scientific Papers in IGO Documents.

Direct citation of scientific papers by IGOs is rare: Only about 0.71% of Scopus-indexed papers published since 2015 are cited in IGO documents. Despite their rarity, these papers are critical entry points connecting scientific knowledge to IGO advisory documents, as most scientific papers maintain indirect connections to IGO documents within three to four citation steps (SI Appendix, Fig. S1). Rates at which scientific papers are cited in IGO documents exhibit substantial geographic disparities: Western Europe (particularly Nordic countries), Australia, New Zealand, and South Africa exhibit the highest rates, whereas the rates of Eastern Europe and East Asia remain below global averages (Fig. 1A, first column; SI Appendix, Table S1). This concentration is also observed at the institutional level: Some European universities (e.g., Wageningen University) rank highly in citations in IGO documents despite modest traditional rankings (SI Appendix, Table S2). This geographic concentration reflects well-established pro-Western biases in policy-relevant knowledge production (24, 30). To understand these disparities, we examined three lagged predictors of citation in IGO documents (Methods), which were all measured from papers published before the focal paper: publication journal (J), IGO-cited references (R), and coauthors’ prior IGO-cited papers (A).

Fig. 1.

Academic pathways to IGO documents: citation distance and uptake predictors. (A) Ratio of observed to predicted papers cited in IGO documents for the 50 most productive countries as predictors are added sequentially: baseline count (N) plus journal (J), shown as N + J, plus the number of IGO-cited references (R), shown as N + J + R, and plus authors’ prior papers cited in IGO documents (A), shown as N + J + R + A. (B) Rates of citation in IGO documents by journal volume: low-volume outlets ($\le$500 papers in 2020, Left) versus high-volume outlets (Right). (C) Probability of citation in IGO documents (blue line, Left axis) and paper count (gray bars, Right axis) by number of IGO-cited references. (D) Same as (C) but for a maximum number of authors’ prior papers cited in IGO documents.

The first two predictors, journal venue and the prior IGO-cited references, can be interpreted as markers of salience, which refers to how research is positioned within policy-relevant channels. The journal venue predictor explains cross-national disparities in IGO citation rates (i.e., the share of papers cited by IGOs) (Fig. 1A, second column). Some low-volume journals show high IGO citation rates; for example, the Quarterly Journal of Economics (67.5%). High-volume journals such as Environmental Research Letters (24%), Nature (6.9%), Science (4.2%), and Proceedings of the National Academy of Sciences (PNAS) (4.5%) are also high-probability sources of IGO-cited papers. These papers also exhibit higher academic impact metrics (SI Appendix, Fig. S2), consistent with prior evidence of science–policy coevolution (22). Citation proximity to prior IGO-cited paper shows strong associations with subsequent uptake: Papers with no IGO-cited references show 0.25% uptake probability, whereas those referencing eight or more IGO-cited papers approach 30% (Fig. 1C). These salience-related factors jointly account for variance in international patterns of science–IGO translation (Fig. 1A, third column).

The third predictor—prior citations of coauthors in IGO documents—captures a distinct dimension: credibility attributed to specific scientists. Probability of citation in IGO documents exceeds 10% when any coauthor has eight or more papers cited in IGO documents (Fig. 1D). Including this author-level predictor in models reduces residual variance, with 36 of 50 leading countries converging within $\pm 10\%$ (Fig. 1A, rightmost column). Notable exceptions include countries facing geopolitical tensions, such as Ukraine and Taiwan, whereas East Asian countries (China, Japan, and South Korea) show marginal underrepresentation, which is not explained by our current metrics (SI Appendix, Table S1). Whereas salience operates through research positioning, credibility attaches to individuals—a distinction central to understanding concentration at the science–IGO interface.

Structural Patterns of Citation Concentration in IGO Documents.

Given indications that scientists’ credibility may play an important role in determining what science is cited in IGO documents, we investigated how scientific knowledge flows into IGO documents, focusing on mediator scientists. To map the concentration of such citations among scientists across domains, we classified 230,737 IGO-cited papers into 23 domains through citation network clustering (Methods; SI Appendix, Table S3). The results (Fig. 2A) revealed diverse scientific knowledge in IGO documents, with rapid expansion in a few notable areas: COVID-19 research, circular economy, and data science & AI (SI Appendix, Fig. S3).

Fig. 2.

Concentration and collaboration among Highly IGO-Cited Scientists (HIC-Sci). (A) Citation network visualization of papers cited in IGO documents, with colors indicating 23 domains. (Layout calculated using ForceAtlas2.) (B) Log–log complementary cumulative distribution of papers cited in IGO documents per scientist for representative domains. (C) Author concentration and temporal stability. Scientists are ranked by the number of their papers cited in IGO documents. The y-axis shows the HIC-Sci share—the proportion of top-ranked scientists accounting for 30% of papers cited in IGO documents (lower values $=$ stronger concentration). Left: Concentration across the full period. Center: Three subperiods with equal numbers of IGO documents; lines connect each domain across periods. Right: Same metric applied to journal publications in general journals, specialized journals, and other collections. (D) Normalized rich-club coefficient measuring the tendency of top scientists cited in IGO documents to collaborate more than expected. (E) Coauthorship networks of 300 top scientists cited in IGO documents in two representative domains (climate modeling and data science & AI). (F) Domain-level comparison plotting the proportion of the HIC-Sci (30% of IGO-cited papers) against largest connected component size among the top 300 scientists cited in IGO documents, revealing field-specific concentration patterns across all 23 domains. (G) Relationship between IGO document engagement recency and research homogeneity across fields. The scatter plot shows the correlation between the recent IGO document ratio (x-axis) and the average similarity of papers cited in IGO documents (y-axis) across research domains. Point colors indicate the proportion of top scientists accounting for 30% of IGO-cited papers, ranging from red to blue.

Citation to scientific papers in IGO documents is highly concentrated in every domain, but the degree of concentration varies substantially across domains. The complementary cumulative distribution functions (CCDFs) of the number of IGO-cited papers per scientist, shown in Fig. 2B, exhibit heavy-tailed distributions across all clusters, with statistical tests rejecting exponential decay (SI Appendix, Table S4). Although log-normal fits are statistically superior to pure power laws—a frequent observation in the empirical literature (31)—the power law exponent $\alpha$ of the tail distribution expresses the intensity of concentration. Climate modeling ($\alpha =2.18$) and marine conservation ($\alpha =2.49$) demonstrated high concentration, while data science & AI ($\alpha =4.06$) demonstrated more distributed citation patterns.

For comparison, we defined “Highly IGO-Cited Scientists” (HIC-Sci) as the smallest set of authors, ranked by publication count, whose cumulative output accounts for 30% of IGO-cited papers within a given domain. HIC-Sci represented $<$5.0% of IGO-cited authors in most domains (Fig. 2C, Left), highlighting disproportionate bridging roles. Domain asymmetry reflected patterns in the CCDF. Climate modeling (0.83%; 300 individuals) accounted for 30% of IGO-cited papers, with similar proportions for animal diseases (0.65%) and tobacco control (0.84%). By contrast, HIC-Sci fractions were approximately five times higher in data science & AI (4.4%) and circular economy (4.0%), suggesting more distributed citation patterns.

These concentration levels are equal to or exceed those observed in scholarly publishing. In Fig. 2C (Right), many domains appear below top general journals (e.g., Nature, Science, PNAS, Nat. Commun., and Sci. Adv.), indicating that a smaller share of authors is sufficient to generate 30% of IGO-cited papers. Climate modeling is among the most extreme in concentration levels, rivaling or surpassing top specialty journals (e.g., JAMA, Cell, Lancet, Nat. Phys., and Nat. Chem.). Separately, a large fraction of domains exhibited higher concentrations than three comparison baselines: PLOS ONE, the full corpus of all papers, and the top 0.71% most-cited papers (a threshold matching the IGO citation rate). This concentration pattern was confirmed using the Gini coefficient (SI Appendix, Fig. S4) as an alternative metric. Together with the heavy-tailed CCDFs (Fig. 2B), these comparisons indicate that citation by IGOs follows cumulative-advantage dynamics akin to high-impact publishing.

Crucially, this concentration is temporally persistent. Fig. 2C (Center) partitions the corpus into three periods with equal numbers of IGO documents. For each domain, the figure plots the fraction of HIC-Sci across the periods, with lines connecting period-specific values. Although a few domains showed minor fluctuations between periods, the general trend was temporal consistency. Thus, despite inclusivity initiatives undertaken by IGOs such as the IPCC, WHO, and IPBES and efforts to increase the diversity of scientific uptake in IGO documents, concentration shows little change over time.

What structural mechanisms might sustain such persistent concentration? Coauthorship structures provide evidence for our second hypothesis–relational concentration. All 23 domains exhibited normalized rich-club coefficients [a metric quantifying whether highly connected nodes preferentially connect (32)] exceeding unity (Fig. 2D), indicating preferential HIC-Sci collaboration. Such patterns have been documented in invisible colleges (27) and IPCC networks (24). However, because rich-club coefficients are not directly comparable across networks with different sizes and densities, we further examined the actual connectivity among the top 300 IGO-cited scientists in each domain.

Network visualizations reveal distinct topologies (Fig. 2E; see SI Appendix, Fig. S5 for additional domains): The climate modeling network resembles a dense, single-component network, while data science & AI reflects dispersed interconnectivity with no central synchronization. This qualitative contrast is quantified in Fig. 2F. Mature environmental domains—climate modeling, marine conservation, and environmental conservation—cluster in the lower-right region, where concentration is high and large connected components among top scientists are present. In contrast, emerging fields such as data science & AI and circular economy occupy the upper-left, exhibiting lower concentration among top scientists.

These structural variations correspond to domain characteristics (Fig. 2G). The recency of IGO document engagement (%IGO-Docs After 2018, x-axis) proxies domain maturity. The y-axis measures paper textual similarity, indicating conceptual coherence. Color encodes concentration, where warmer hues indicate fewer scientists generating 30% of IGO-cited papers. The COVID-19 pandemic and data science & AI domains showed recent engagement, low similarity, and distributed citation patterns. Tobacco control, animal diseases, and climate modeling displayed established engagement, high similarity, and concentrated citation patterns. These patterns suggest that domain maturity and increasing conceptual coherence are associated with stronger citation concentration in IGO documents and denser collaboration networks (11).

Relational and Cognitive Pathways of HIC-Sci.

To understand the mechanisms underlying HIC-Sci’s disproportionate citation rates in IGO documents, we examined both the speed and nature of their uptake in IGO documents. As shown in Fig. 3, HIC-Sci papers had substantially higher rates of international coauthorship than other papers cited in IGO documents across nearly every domain. The difference was most pronounced in fields such as climate modeling, vector-borne diseases, and environmental conservation, where HIC-Sci papers consistently exceeded both the average level of collaboration among IGO-cited papers and the overall Scopus baseline since 2015 (orange vertical line).

Fig. 3.

International coauthorship across scientific domains. Blue crosses denote HIC-Sci papers, and gray crosses denote other papers cited in IGO documents. The orange vertical line indicates the overall proportion of international coauthorship in Scopus-indexed publications (2015–2023).

This elevated international collaboration is associated with broader reach across IGO documents, although the geographic distribution of HIC-Sci networks exhibits distinct regional asymmetries. Fig. 4 presents a dense coauthorship nexus in Europe, where node sizes correspond to the HIC-Sci papers cited in IGO documents of each country, and edges link country pairs whose Jaccard index of coauthored HIC-Sci papers exceeds 0.065. The figure also presents a sensitivity analysis demonstrating robustness across 5% and 8% thresholds (SI Appendix, Fig. S6). This configuration reflects a center–periphery structure long identified in the global environmental assessment and policy-relevant knowledge production (3), which has also been shown to characterize geographically uneven patterns of scientific output in policy-adjacent research fields (30). These findings demonstrate that while the overall network of research cited in IGO documents is international in scope, citation concentration remains unevenly distributed across regions and institutions.

Fig. 4.

Country-level coauthorship network among HIC-Sci. The node size reflects the number of HIC-Sci papers cited in IGO documents from each country. Edges connect country pairs whose Jaccard index of coauthored HIC-Sci papers exceeds 0.065.

Fig. 5 provides evidence for our third hypothesis: cognitive concentration. The systematically shorter citation lags for HIC-Sci across nearly all domains demonstrate close temporal alignment between their research agendas and IGO priorities. This temporal advantage, along with their focus on policy-aligned vocabularies (Table 1) is consistent with processes of epistemic selection, in which standardization and isomorphism (33, 34) may favor certain knowledge frameworks, with consequences for methodological diversity (35).

Fig. 5.

Citation lag in IGO documents by research domain. Gray bars show the average years between publication and citation in IGO documents for other papers cited in IGO documents. Blue bars show the same metric for HIC-Sci papers. With data coverage from 2015 to 2023, estimated lags are shorter than those obtained from analyses that use longer historical windows.

Table 1.

Representative terms by HIC-Sci Papers Rate (HPR)

HPR	Representative terms
$\mathrm{HPR}=0$	Climatic importance, discussion issue meeting, discussion meeting, ecological function, findings research,
	space station, soil use, bb biomass burning, stations synoptic, cycle mechanism pattern
$0<\mathrm{HPR}<0.3$	Adaptive measure, eof analysis function, eof analysis, double low srccombining, issue meeting, double
	srccombining, estimate surface, convection-permitting scale, terrestrial tropical, carbon estimate
$0.3\le \mathrm{HPR}<1$	Era5 reanalysis, adjustment bias, cfc-11 emission, cmip6 simulation, change profound, Canadian earth,
	Canadian earth system, cmip6 multimodel, action adaptation, gwls level warming, (*scenario ssp5-8.5,
	*ssp1-2.6, *ssp2-4.5, *high-emissions scenario)
$\mathrm{HPR}=1$	Cmip6 ensemble, feedback land–atmosphere, ensemble geoengineering large, clams model, geoengineering
	large, mixed-layer model ocean, hazard indicator, hazard risk, sandy shoreline, model previous version

Top 10 most frequent and representative terms for each HPR category, extracted from the titles and abstracts of papers published since 2018 to focus on recent, specialized terminology. Terms are generally listed in descending order of frequency, but those marked with an asterisk (*) are included for their distinctiveness despite their lower frequencies. HPR indicates the proportion of total occurrences of a term found in HIC-Sci papers, which constitute 30.0% of all papers analyzed (Methods).

Measuring IGO Citation Patterns of HIC-Sci in Climate Modeling.

To illustrate how concentration patterns manifest in specific domains, we examined climate modeling, a domain in which IGO citation concentration is exceptionally high. We used authorship of IPCC reports as a proxy for scientists’ engagement with IGO assessment processes. Fig. 6 shows that approximately 30% of the top 100 IGO-cited scientists in climate modeling have served as IPCC authors. Among all 300 HIC-Sci in climate modeling (those accounting for 30% of IGO-cited papers), 93 (31%) contributed to IPCC reports, confirming a strong association between IPCC authorship and IGO citation concentration. The concentration of IPCC authorship among HIC-Sci shows patterns consistent with cumulative advantage, where initial recognition may be associated with subsequent opportunities for engagement with IGO assessment processes.

Fig. 6.

Scientific advisory participation rates by IGO citation rank in Climate Modeling. Proportion of climate modeling scientists cited in IGO documents ($n=36,143$) serving as IPCC report authors, binned by number of papers cited in IGO documents.

To assess the association between HIC-Sci authorship and IGO citation rates while partially controlling for research content, we employed cosine similarity–based semantic matching (Methods) to pair each HIC-Sci-authored paper with a topically comparable non–HIC-Sci paper. We found that HIC-Sci papers had higher IGO citation rates compared to matched non–HIC-Sci papers (50.1% [1,228/2,450] versus 28.1% [688/2,450]). This gap suggests that IGO citation rates are not attributable solely to the content of the research itself.

The terminology analysis revealed that HIC-Sci authors use more policy-ready modeling terms, while non–HIC-Sci authors use more observational terms. We calculated the HIC-Sci papers share rate (HPR), defined as the proportion of occurrences of a term in papers with at least one HIC-Sci author. Terms with HPR $=$ 0 (appearing only in non–HIC-Sci papers) were primarily observational, such as “bb biomass burning,” “ecological function,” and “soil use” (Table 1). By contrast, terms with HPR approaching or equal to 1 (predominantly in HIC-Sci papers) contained policy-ready modeling vocabularies, such as “cmip6 simulation,” “ssp scenarios,” “hazard risk,” and “geoengineering large ensemble.” This lexical specialization, combined with shorter citation lags (Fig. 5), suggests a pattern of cognitive concentration: HIC-Sci research is more frequently articulated in policy-compatible formats through standardized modeling frameworks, while non–HIC-Sci research maintains its focus on empirical foundations.

Synchronization in IGOs’ Citations of Scientific Papers.

The concentration of citations among HIC-Sci extends beyond citation patterns to the institutional pathways through which their research diffuses. HIC-Sci papers circulate across a broader range of IGOs than those authored by other scientists. Fig. 7 shows that HIC-Sci papers were less frequently cited by single IGOs and more frequently cited by multiple IGOs. HIC-Sci papers reached 1.73 times as many IGOs as non–HIC-Sci papers on average, a pattern consistent across domains. This multi-IGO “source-node” position is consistent with cumulative advantage mechanisms in epistemic communities and means that HIC-Sci are capable of seeding coordinated science uptake across IGOs (36).

Fig. 7.

Distribution of number of IGOs citing papers. Comparison of papers authored by Highly IGO-Cited Scientists (HIC-Sci, green) and other papers (non–HIC-Sci, blue).

The aggregate cocitation network (Fig. 8A) revealed a temporal hierarchy in scientific knowledge diffusion among IGOs. Arrows indicate directionality from early- to later-citing organizations, with major multilateral institutions such as the United Nations (UN), the World Bank, and the WHO showing the highest frequencies as initial citers of emerging scientific evidence. Small specialized agencies are likely to cite papers later, positioning major IGOs as centers of knowledge-gathering within the policy system. Such synchronization, in which later-citing agencies converge on sources already validated by major IGOs, amplifies the citation concentration of HIC-Sci.

Fig. 8.

Cocitation networks of IGOs. (A) Aggregate network across all domains. Node size is proportional to the number of citing papers. Edge width indicates citation-set overlap (cocited papers divided by a smaller citation set). Only edges with overlap $>$0.3 are shown. Arrows point from early- to later-citing IGOs. (B–E) Domain-specific networks. (B) environmental conservation, (C) climate modeling, (D) COVID-19 pandemic, (E) data science & AI.

Domain specific cocitation networks (Fig. 8B–E) revealed field-specific variation in institutional cascades. Climate modeling showed a tightly interconnected core around the IPCC and UN, with a citation overlap of 0.59, suggesting limited source diversity among leading climate-focused IGOs. Environmental conservation followed the same pattern, retaining the UN-FAO centrality. The COVID-19 pandemic network showed high cohesion, with the WHO at its center to connect multiple public health IGOs, similar to what happens during rapid instrumental knowledge mobilization. Data science & AI showed a fragmented topology without major nodes.

Discussion.

The uptake of scientific knowledge by IGOs is highly skewed toward a small, tightly connected group of scientists we refer to as HIC-Sci. This concentration reflects how IGOs manage information abundance under time constraints (16). Across the 23 domains examined in this study, 0.7% to 4.4% of authors accounted for 30% of IGO-cited papers, with concentration levels comparable to or exceeding those observed in leading academic journals (Fig. 2C). Papers authored by scientists with prior IGO-cited papers are substantially more likely to be cited in IGO documents (Fig. 1D), a pattern consistent with cumulative advantage dynamics (7). In climate modeling, HIC-Sci papers are cited in IGO documents at significantly higher rates than non–HIC-Sci papers even after controlling for topical similarity (50.1% versus 28.1%). Collectively, the findings suggest that the credibility levels of authors—in addition to the research content itself—are systematically related to incorporation in IGO documents.

This structural concentration pattern is highly correlated with relational positions and common cognitive approaches. For all 23 domains, HIC-Sci coauthorship networks had normalized rich-club coefficients exceeding 1 (Fig. 2D), suggesting that top scientists cited in IGO documents preferentially collaborate with one another. In climate modeling, 31% of HIC-Sci (93 out of 300) had served as authors of IPCC assessment reports (Fig. 6), indicating dense interconnections between HIC-Sci and IGOs producing advisory documents. This embeddedness is also temporal: HIC-Sci papers maintained consistently shorter lags from publication to citation in IGO documents across almost all domains (Fig. 5). In addition to network position, HIC-Sci exhibited distinct cognitive orientations, using relatively standardized, policy-oriented vocabularies (e.g., “CMIP6 simulation” and “SSP scenarios”), while non–HIC-Sci papers used observational terminology (e.g., “ecological function” and “soil use;” Table 1). This pattern is consistent with the concept of epistemic communities (36), whose shared causal beliefs and policy-relevant vocabularies make knowledge more actionable.

While this reliance on trusted interlocutors enhances efficiency, it raises questions of legitimacy. One dimension of this tension is geographic. Although HIC-Sci papers exhibited high levels of international coauthorship (Fig. 3), their distribution followed a core–periphery structure centered on the European Union (Fig. 4). EU countries and the United Kingdom showed higher probabilities of citation in IGO documents (Fig. 1A), differences that are partially—but not fully—accounted for by author-level credibility (Fig. 1D). These patterns are consistent with prior evidence of spatial inequality in policy-relevant knowledge production (24, 30).

Concentration patterns were also temporally stable. Dividing the study period into three segments revealed little systematic change over time (Fig. 2C), which is consistent with theories of institutional path dependence (14, 15). This persistence is remarkable given IGOs’ continued efforts to diversify their regional and disciplinary representation, implement conflict-of-interest protocols (20), and standardize evidence-integration procedures (21). The contrast between this persistence and these reform efforts suggests that structural citation patterns may operate on longer time scales than procedural reforms and that broadening established knowledge flows might require careful institutional arrangements.

At the same time, concentration varies substantially across domains. In mature fields, higher concentration appears to support efficient knowledge transmission. As we see in climate modeling, where only 0.83% of authors account for 30% of IGO-cited papers, dense collaborations and cocitation networks centered around the IPCC allow for the synthesis of complex evidence into authoritative policy advice (37, 38). Concentration does not imply exclusion. Non–HIC-Sci scientists contribute diverse observational and diagnostic knowledge that HIC-Sci integrate into policy-relevant frameworks (Table 1). In this respect, the IPCC acts as a hub of scientific advice to policymakers that organizes information inflows and outflows and promotes comparability among studies, while preserving epistemic diversity (39).

By contrast, data science & AI showed uptake of knowledge from a broader distribution of scientists and no single dominant IGO hub citing the work of scientists (40). AI ethics frameworks show convergence at the level of abstract principles (41) but divergence in interpretation and implementation across contexts (42). Aligning with proposals for IPCC-like advisory bodies in AI governance (40, 43), the United Nations General Assembly established the Independent International Scientific Panel on Artificial Intelligence in August 2025 (44). This institutionalization may precede, rather than follow, epistemic convergence, raising distinct legitimacy challenges. In climate science, advisory authority derives from well-established scientific consensus (45). AI governance, by contrast, involves deeper normative disagreement over fundamental concepts such as fairness (46). In such settings, legitimacy concerns feature prominently in national AI governance strategies (47), and prior work suggests that transparency, inclusive representation, and procedural safeguards matter for maintaining legitimacy when concentration precedes epistemic stabilization (48). Therefore, for emerging domains such as AI governance, systematic monitoring of citation patterns and transparent reporting of advisory participation may be important safeguards for procedural legitimacy.

Existing IGO practices provide partial precedent for this level of transparency. The IPCC has been explicit in specifying the criteria for expert selection and literature evaluation (49). The WHO has published standards for evidence appraisal and conflict-of-interest management (20), and the IPBES has formalized protocols for integrating Indigenous and local knowledge (50). However, these initiatives focus on the transparency of inputs and processes rather than outcomes; efforts to track whose knowledge is ultimately incorporated into policy remain limited. Beyond disclosing which scientific literature is cited by IGOs in their policy advisory documents, author-level analyses such as ours can reveal who is cited in them and who is not, making the concentration of scientific knowledge more visible and enhancing accountability for knowledge selection. Balancing credibility, salience, and legitimacy in science–policy interfaces remains challenging (3), but improving the visibility of knowledge-selection processes, particularly at the outcome level, may help address this tension.

This study has several limitations. The citation data captured outcomes of knowledge selection but did not reveal knowledge-creation processes. We observed citation concentration in IGO documents among HIC-Sci but could not disentangle whether higher citation rates reflect expertise, institutional access, or both. Future research on scientists’ career trajectories and IGO affiliations could clarify how scientific knowledge is selected or excluded. Extending this framework to national-level policy documents would also help assess whether similar concentration patterns emerge outside of intergovernmental settings. A long-term analysis of historical citation patterns could also illuminate how IGOs’ efforts to broaden participation shape these outcomes over time.

Methods

Dataset.

We used two data sources: Overton and Scopus. Overton is a comprehensive database of policy documents and their citations of scientific papers, whereas Scopus was chosen for its extensive coverage and accurate author identification (51). Our analysis focused on papers cited in IGO documents and was restricted to papers published since 2015 to capture recent trends in science–IGO interactions.

From Overton, we initially retrieved 305,635 papers published since 2015 that were cited in IGO documents. We then matched these Overton records with Scopus entries using DOI. After this process, the final dataset comprised 230,737 papers that could be definitively linked between Overton and Scopus. While the Overton database has a bias toward English literature, its coverage of IGO documents is highly reliable for several reasons. First, IGO documents exhibit a citation density of academic papers several times higher than that of government documents. Second, major IGOs such as the WHO and UNESCO systematically publish their documents online. Finally, the majority of international organization documents are published in English, resulting in high coverage and processing accuracy (23).

Measuring Distance to IGO Documents.

Following the concept of paper-patent distance (52), we defined each paper’s distance to IGO documents as the shortest undirected path connecting it to an IGO document in a combined citation network of both scientific papers and IGO documents. In this network, each node represents either a paper or an IGO document, and edges represent citation relationships that can be traversed in both directions. Papers directly cited by an IGO document have a distance of 1. For any other paper, the distance is 1 plus the shortest path to any distance 1 paper, regardless of the citation direction. For instance, a paper with distance 2 may either cite or be cited by a distance 1 paper but is not directly cited by IGOs.

Estimating the Number of Papers Cited in IGO Documents for Each Country.

To estimate how many papers from each country were cited in IGO documents, we developed a sequential prediction approach. Starting from a baseline model, we incrementally added three lagged predictors—each measured from papers published before the focal paper—to enhance prediction accuracy:

• Baseline (N): The simplest model considers only the total number of papers published by each country, treating all papers equally without distinguishing among different characteristics.

• N + J (Adding journal venue): This model improves accuracy by incorporating publication journal as a predictor. We grouped papers by journal and estimated the probability of citation in IGO documents based on historical citation rates for each journal.

• N + J + R (Adding IGO-cited references): This model further refines estimates by considering each paper’s reference list. Papers are grouped according to how many of their references have previously been cited in IGO documents. To handle wide ranges, we grouped these counts into logarithmic bins (e.g., 0, 1, 2 to 3, 4 to 7 references, and so forth).

• N + J + R + A (Adding authors’ prior IGO-cited papers): The most comprehensive model also incorporates authors’ past publication histories, using logarithmic bins of the maximum number of prior papers cited in IGO documents among coauthors.

For each model specification, we used papers published before our estimation period to calculate historical probabilities of citation in IGO documents in each group. Specifically, for a group $g$ containing $n_g$ total papers, of which $m_g$ were cited in IGO documents, the probability was computed as $P(\text{IGO-cited}\mid g)=m_g/n_g$.

To estimate the number of papers cited in IGO documents for a particular country during the estimation period, we summed these probabilities across all papers from that country. Formally, if paper $i$ belonged to country $x$ and was assigned to group $g_i$, the estimated total number of papers cited in IGO documents for country $x$ was $\sum _{i\in C_x}P(\text{IGO-cited}\mid g_i)$, where $C_x$ denotes the set of all papers affiliated with country $x$. Comparing observed counts with these estimates revealed which countries exceeded or fell short of predictions, as each predictor was added (Fig. 1A).

Grouping Papers Cited in IGO Documents by Research Domain.

To identify major research themes among the papers cited in IGO documents, we constructed a citation network of these papers and applied the Leiden clustering method (53) to detect communities of frequently cociting papers. While the Louvain algorithm (54) has been widely adopted for community detection, we employed the more recent Leiden algorithm, as it guarantees connected communities and overcomes the resolution limit problem inherent in Louvain’s greedy optimization. This capability proved particularly crucial for detecting smaller, specialized research communities within our large-scale network of over 230,000 papers, where niche but policy-relevant areas might otherwise be absorbed into larger clusters.

After cluster identification, we characterized each domain through a term frequency–inverse document frequency (TF-IDF) analysis of paper abstracts. TF-IDF effectively balances term frequency with document specificity, revealing cluster-distinctive vocabulary while maintaining interpretable term weights—an advantage over embedding-based methods when methodological transparency is essential. We then employed GPT-4o to synthesize the highest-scoring terms into concise thematic labels, providing clear descriptors for each research domain.

Identifying HIC-Sci and Comparing the Concentration with Journal Publications.

Within each research domain, we identified scientists whose research is frequently cited in IGO documents—termed HIC-Sci—using a concentration-based approach similar to that of Larivière et al. (6). First, we considered each paper cited in IGO documents within a domain, excluding papers with more than 30 authors, for which individual contributions become difficult to discern. We then calculated each scientist’s total contribution across all their papers cited in IGO documents within the domain. Next, we ranked the scientists in descending order based on these summed contributions. Finally, we selected the top scientists whose combined contributions accounted for 30% of all papers cited in IGO documents within that domain.

To validate our concentration analysis against broader publication patterns, we compared author concentration in papers cited in IGO documents with that in general scientific literature. For journals and collections exceeding 10,000 papers, we randomly sampled 10,000 papers to ensure computational feasibility. This 0.71% sampling rate matched the proportion of papers cited in IGO documents in the overall scientific corpus, enabling direct comparison of concentration patterns between IGO-cited and general scientific outputs. Our analysis revealed that the concentration of authorship in papers cited in IGO documents followed patterns similar to those observed in prestigious journals, which suggests that citation in IGO documents is structured through selection and cumulative advantage processes analogous to those operating in high-impact academic publishing.

Textual Similarity Analysis in Each Research Domain.

To quantify the conceptual coherence within a research domain, we measured the average pairwise textual similarity among papers cited in IGO documents in the domain. We first generated embedding vectors for each paper using OpenAI’s text-embedding-3-small model, which produces 1,536-dimensional dense vector representations that capture semantic meaning from paper abstracts and titles. This embedding model was chosen for its ability to encode nuanced scientific concepts while maintaining computational efficiency for large-scale analysis.

Matching Semantically Similar HIC-Sci and Non–HIC-Sci Papers.

To assess how HIC-Sci authorship relates to citation rates in IGO documents in climate change research, we used Elsevier’s International Centre for the Study of Research (ICSR) dataset, which contains metadata on more than 90 million scholarly papers indexed in the Scopus database. The ICSR dataset includes Elsevier’s United Nations Sustainable Development Goals (SDGs) research classifications, which categorize research papers as contributing to or addressing one of the SDGs (55). Using the SDG categories, we identified 338,250 research papers related to SDG 13: Climate Action published after 2016, of which 5,167 had at least one HIC-Sci author.

For each HIC-Sci–authored paper in the dataset, we searched for semantically similar papers across 20 million Scopus papers published after 2016 and before May 2023. The search used miniLM (56)-generated abstract embeddings and elastic KNN (57) to find the top 10 nearest neighbors with a cosine similarity score higher than 0.8, a threshold validated manually in previous research (58). The matches were filtered to include pairs where both papers were published no more than one year apart and had entirely different author teams, and where one paper had at least one HIC-Sci coauthor while the other had no HIC-Sci authors at all. This resulted in a final dataset of 2,450 pairs, with some HIC-Sci authored papers having multiple matches, which we treat as independent pairs.

Author Matching Between IPCC Members and HIC-Sci.

We compiled author lists from 12 IPCC assessments and special reports published between 2011 and 2023 (detailed in SI Appendix, Table S5). Names were extracted from the PDF appendices, converted to lowercase in a standardized “surname, given name” format, and deduplicated, resulting in a list of 1,639 unique authors.

To identify overlaps with the 36,143 IGO-cited scientists in the climate modeling domain, including the 300 HIC-Sci, we applied fuzzy string matching using the token_set_ratio function from the FuzzyWuzzy library. For each IPCC author, we identified the best-matching name in the list of scientists based on the token set ratio score. Matches with a score higher than 83 were considered valid. This threshold was selected to allow for minor name variations while minimizing false positives. In total, 824 scientists, including 93 HIC-Sci, were identified as authors of at least one of the selected IPCC reports.

Identifying Emerging Terms in Climate Modeling Papers.

We analyzed recent trends in climate modeling research by identifying new scientific terms introduced from 2018 onward. Although our dataset started from 2015, we began the analysis from 2018 to allow for a sufficient buffer period to accurately capture emerging terminology.

We extracted two- and three-word phrases (bigrams and trigrams) from abstracts of climate modeling papers published from 2018 onward, applying lemmatization to standardize terms to their base forms. Papers were categorized into two groups: HIC-Sci papers with at least one HIC-Sci coauthor and non–HIC-Sci papers without any HIC-Sci coauthors. For each phrase, we calculated its HIC-Sci papers share rate (HPR)—the proportion of occurrences in HIC-Sci papers relative to the total. Terms were classified by HPR: exclusive to non–HIC-Sci ($\mathrm{HPR}=0$), underrepresented in HIC-Sci ($0<\mathrm{HPR}<0.3$), overrepresented in HIC-Sci ($0.3\le \mathrm{HPR}<1$), or exclusive to HIC-Sci ($\mathrm{HPR}=1$).

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Raw data from Overton and Scopus are proprietary and cannot be redistributed due to licensing restrictions. All analysis code, search procedures, and data processing steps are documented in a public repository: https://github.com/asatani/IGO_science_citation (59). Researchers wishing to reproduce the study can obtain access to Overton (https://www.overton.io/) (60) and Scopus (https://www.scopus.com/) (61) directly under their respective licensing terms and can then use the public code to reproduce the analyses.