The proximal cause of redundant synthesis is the independent formulation of near-identical questions by multiple teams, often without adequate scoping of the existing evidence landscape. In a meta-analysis of 21 meta-research studies representing 3621 original studies and protocols, the proportion of new clinical studies justified using a systematic review ranged from 16% to 87%, with a pooled mean of only 42% (95% CI 36%-48%) [17]; a systematic examination of randomized controlled trials likewise found that most reports did not cite prior research relevant to the trial question [38]. Justification using existing reviews therefore remains inconsistent [17], even though the evidence-based research approach is widely endorsed in principle [19,39].

The PICO (population, intervention, comparison, and outcome) framework, while essential for defining precise questions, can mask substantial overlap: teams addressing what they believe to be distinct questions may investigate nearly identical populations and interventions, and because overlap is conventionally assessed element by element rather than at the level of the question, partial congruence goes unrecognized. Among overlapping network meta-analyses, a median of 73% of nodes, 79% of treatments, and 48% of references were shared between index and overlapping analyses [8]. Lund and colleagues [18] identify the lack of systematic evaluation of existing evidence as a primary contributor to redundancy in health research. This structural gap concerns the absence of mandated pre–protocol scoping requirements rather than any claim that no scoping occurs: many researchers establish the scope of the research, but without a standard requirement, quality is variable and unenforceable.

PROSPERO was established to enhance transparency and reduce unintended duplication by requiring the deposit of a protocol before work begins [14]. Registrations rose 10-fold from 63 per month in 2012 to about 800 per month in 2017, exceeding 26,500 records [14], and prospectively registered reviews achieved higher AMSTAR (A Measurement Tool to Assess Systematic Reviews), R-AMSTAR (revised AMSTAR), and PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) scores than nonregistered reviews [32]. At least 5 registration options now exist, including INPLASY (International Platform of Registered Systematic Review and Meta-analysis Protocols) and the Open Science Framework [40]. Its limitations as a duplication-prevention mechanism, however, are structural rather than incidental.

Beresford and colleagues [13] found substantial duplication among PROSPERO registrations during the COVID-19 pandemic. Of 1054 COVID-19 registrations submitted between March 2020 and January 2021, 138 were submitted when a protocol with a similar PICOS (population, intervention, comparison, outcome, and study design) was already registered; duplication was most pronounced for treatment reviews, with 14 separate registrations addressing hydroxychloroquine efficacy [13]. Surveyed authors mostly reported having identified similar reviews when searching PROSPERO beforehand, citing differences in PICOS or planned analyses, poor quality of previous registrations, and the need to update evidence [13]. The platform relies on manual, user-initiated searching; its newer overlap-flagging features remain author-assessed rather than semantically automated, and detection occurs at registration rather than at topic selection. The mechanism is therefore reactive rather than preventive, compounded by the fact that registration remains far from universal [36].

The persistent production of redundant syntheses cannot be explained by methodological lapses alone. Incentive structures rewarding publication volume, journal prestige, and citation counts favor output irrespective of genuine knowledge gaps [1,2], and competition for positions and grants privileges productivity metrics over epistemic value [41]. Systematic reviews suit this environment: they need no primary data collection, fit doctoral or grant cycles, and are well cited [2,42]; many function primarily as easily produced publishable units, a type readily serving sponsors' marketing interests [2]. The result is overproduction on popular topics, misaligning synthesis effort with evidence need [15]. Thousands of research groups, guideline bodies, and funders operate with varying awareness of the evidence landscape, and no central mechanism prioritizes topics or redirects resources from saturated questions [3]. Ravaud and colleagues [16] diagnosed this as a core pathology, proposing a transition from a decentralized, investigator-driven model to a coordinated evidence ecosystem.

The COVID-19 pandemic compressed and amplified these dynamics, serving as a natural experiment [43]. The acute demand and rapid accumulation of primary studies pushed review production beyond both the rate of new primary evidence and the ecosystem's absorptive capacity. The acceleration came with a measurable decline in rigor: a review of 686 COVID-19 clinical articles found substantially shorter times to acceptance (median 13 days versus 110 days for historical controls) and lower methodological quality across all study designs [44]. The evidence base was correspondingly unstable, and Brubacher and colleagues [33] found that proliferating overlapping syntheses created confusion among decision-makers about which evidence to trust; this episode is better characterized as a systemic failure of the evidence pipeline than as a succession of individual errors [43].

Review output had already grown steeply before the pandemic, exceeding the annual output of randomized trials by 2017 [4]. During the pandemic, this trajectory steepened while the evidence base remained thin and rapidly changing. In one case study, 38 systematic reviews of remdesivir, a new publication every 9 days, appeared within roughly 14 months; 51% were out-of-date at publication and only 21% had been registered in PROSPERO [10]. A parallel analysis of 346 reviews of COVID-19 treatments found that 75.4% had not been registered at all and that only 29.7% of the registered protocols had been published after at least 17 months [36]. The pandemic thus exposed rather than created a structural vulnerability: where a living product was maintained instead, the same question was addressed with a fraction of the effort, with living guidelines updating within 14 days [45,46].

Production-side redundancy concerns duplicated review questions; execution-level overlap concerns repeated inclusion of the same primary studies across multiple reviews on the same topic [27]. Question-level duplication wastes effort before any results exist and is preventable at registration. Primary study overlap becomes visible only once reviews are published, and its principal statistical consequence is the inflation of apparent sample size and event count, overstating the precision of pooled estimates [27]. Overlap is not in itself bias: it may reflect legitimate differences in question framing and can be a diagnostic asset as much as a hazard [26]—the problem arising when such overlap is substantial and unacknowledged [27].

The primary metric for quantifying overlap is the corrected covered area (CCA): 0% to 5% indicates slight overlap, 6% to 10% moderate, 11% to 15% high, and greater than 15% very high [31]. In the original validation study of 60 overviews, only 32 reported overlap at all and the median CCA was 4.0 [31]. Ou and colleagues [5] confirmed across 11 quantitative studies that 68% of the systematic reviews exhibit overlap, with as many as 76 overlapping reviews on a single topic, and that only 36% referenced previous reviews on the same question. Overlap has since been documented across oncology, cardiovascular medicine, genetic epidemiology, and COVID-19 research [8,9,47]: only 12 of 38 index meta-analyses in genetic epidemiology were unambiguously unique, with a mean of 2.6 duplicates per index analysis [9], and multiple cardiovascular oncology reviews of the same drug-class associations showed significant overlap in included primary studies [47]. Clinicians and guideline developers may therefore encounter summaries that appear independent but derive from largely the same data.

Execution-level overlap distorts the perceived weight of clinical evidence. When the same primary studies contribute effect estimates to multiple meta-analyses, apparent consistency across syntheses is often read as replication and robustness, when it frequently reflects shared data rather than independent corroboration [27]. Two caveats are essential. First, this applies specifically to tertiary syntheses such as overviews and umbrella reviews, in which overlapping reviews are themselves pooled; for stand-alone reviews, the primary harms are research waste, confusing guidance, and inefficient resource allocation rather than statistical distortion per se [26,27]. Second, the phenomenon is not confined to overlap between reviews: covert duplicate publication of primary data has repeatedly been shown to inflate meta-analytic effect estimates—the classic case of ondansetron demonstrating that duplicated trial reports can materially change pooled results [48].

Guideline panels are particularly vulnerable when they base recommendations on the number of supporting meta-analyses rather than on independent primary studies. In a cross-sectional survey of 684 health care decision-makers, 97.9% reported regularly or occasionally encountering more than one systematic review on a topic of interest and 50.8% reported difficulty identifying the most trustworthy synthesis, with difficulties relating to lack of time (55.2%), varying methodological quality (54.2%), differences in results and conclusions (49.7%), and variation in included studies (44.6%) [20]. Most respondents compared reviews on relevance, quality, and recency of search, and 87% expressed willingness to use an AI-assisted appraisal tool [20]. A perceptual pathway also operates regardless of pooling: multiple reviews on one topic create an impression of evidential richness that may not correspond to genuine depth [5,35].

The resource implications are substantial: a single systematic review typically requires hundreds of person-hours from protocol development through publication, and duplication of this effort sits within a broader pattern of avoidable waste in the production and reporting of research evidence [6]. Global estimates of the resulting waste nonetheless remain methodologically fragile, resting on extrapolations from small cost-per-review samples and untested counterfactual assumptions; 3 decades after the first formal articulation of avoidable waste, the problem remains unresolved rather than diminishing [6], and the framing has since been extended beyond medicine [49]. A less tangible but arguably more consequential harm is the erosion of trust in systematic reviews as a method [2,50]: conflicting reviews on one topic invite the inference that the methodology is unreliable, even when discordance arises from overlapping but nonidentical inclusion criteria [5,20].

A conceptual challenge intersects all propagation-side concerns: distinguishing legitimate redundancy from wasteful duplication. Some redundancy serves epistemic functions, including methodological triangulation, temporal updating, and contextual adaptation; replication in this sense is a deliberate, design-driven activity whose value depends on what it is intended to establish rather than mere repetition [25]. In practice, these distinctions are hard to operationalize. Rombey and colleagues [51] examined 100 non-Cochrane updates of systematic reviews, published a median of 5 years after the previous version; only half used the term update in the title, none reported using an existing decision tool, and less than one-third changed their conclusion [51]. Currency is more generally neglected: among 147 overviews, only about 1 in 4 considered whether included reviews were up-to-date [21]. CCA captures quantitative overlap but cannot assess the epistemic contribution of a new synthesis [26,31]. The evidence-based research approach provides a principled basis for such judgments [19,39] but has not been operationalized into prospectively applicable criteria, and no normative consensus on acceptable redundancy levels exists.

Because no threshold distinguishes productive from wasteful redundancy quantitatively, the operational question is not how much overlap exists but whether the proposed synthesis could change a decision that the existing evidence base cannot support—by incorporating new primary studies, by addressing a different PICO, or by applying a different analytical method. A new review is justified when its authors can name the decision it would alter and the element of the question that existing syntheses leave unresolved [26,31,35]. Making the criterion explicit also assigns responsibility clearly: investigators would state which decision the synthesis is intended to inform and which element of the question remains unresolved, while reviewers, editors, and funders would adjudicate that claim instead of inferring novelty from its absence.

Current governance mechanisms for preventing redundant synthesis are fragmented and inconsistently enforced. Prospective registration, the primary existing safeguard, suffers from incomplete coverage, delayed publication, and reliance on manual overlap checking [22,36]: three-quarters of published COVID-19 treatment reviews were never registered [36], and quality gains are real but modest [32]. Reporting guidelines compound the problem: PRISMA 2020 specifies how a completed review should be reported but does not require authors to demonstrate that their synthesis addresses an unmet evidence need [37], and the PRIOR statement likewise addresses reporting quality rather than necessity [52]. Kolaski and colleagues [12] articulated principles of good scholarship requiring authors to situate their work within the existing evidence base and justify its contribution, though these remain honored more in the breach. The burden of determining novelty therefore falls on investigators, reviewers, and editors, none of whom have systematic tools for assessing duplication risk, and editorial policies vary widely [3,12].

There is no central mechanism that coordinates synthesis production, prioritizes topics according to evidence needs, or redirects resources from saturated questions. The Future of Evidence Ecosystem series identified this gap and proposed a transition to coordinated, continuously updated synthesis infrastructure oriented to user needs [16,23], and Akl and colleagues [24] proposed Evidence Synthesis 2.0, integrating systematic, scoping, rapid, living, and overview reviews into a coordinated workflow. The COVID-END (COVID-19 Evidence Network to Support Decision-Making) demonstrated both the feasibility of coordination, through shared priority-setting, joint evidence mapping, and collaborative production across more than 50 partner organizations and the difficulty of sustaining it without permanent infrastructure and funding [53]. Frameworks for setting health research priorities have been available for more than a decade, with explicit recommendations that value and waste be assessed [15]; what is missing is not method but mandate.

Governance failures extend to the institutional level. Promotion criteria reward publication volume and citation metrics rather than synthesis quality, coordination, or ecosystem stewardship [1,2], and grant review typically evaluates the intrinsic merit of individual proposals without assessing whether the proposed synthesis duplicates existing work [34]. Hayden and colleagues [34] argued for collaborative systematic review as a model achieving efficiency gains while preserving investigator autonomy, but such models need institutional support that remains underdeveloped [19,39]. The cumulative effect is an ecosystem without feedback: redundant production carries no penalty, and nothing self-corrects. The CCA was proposed in 2014; yet, 10 years later, 68% of the reviews still exhibited overlap and only 9% reported protocol registration [5,31]. The interventions discussed next are therefore evaluated not only on merit but on whether they introduce a feedback signal the system lacks.

Mandating systematic scoping searches as a condition of PROSPERO registration and journal submission is the most immediately actionable intervention, combining high feasibility, a moderate evidence base, high scalability, and moderate risk. Authors would document a systematic search for existing reviews and provide explicit novelty justification referenced to defined PICO elements before registration [19,39]. PROSPERO could require a scoping search and a novelty statement as a condition of registration, and journals could require evidence of preregistration scoping at submission. The PRIOR (Preferred Reporting Items for Overviews of Reviews) statement shows that reporting standards can be extended to a new review type when the field identifies a need [52], and Xu and colleagues [54] found that formal protocol development, including systematic scoping, measurably improves review quality, consistent with the association between registration and higher AMSTAR scores [32]. The main risk is perfunctory searches meeting the formal requirement without genuine assessment [13]; enforcement would require editorial oversight and automated cross-validation against registry content.

The most promising technological intervention is an automated system for detecting whether a proposed review topic already has an adequate evidence base, combining moderate feasibility, an emerging evidence base, high scalability, and significant risk. Machine learning has been applied across screening, data extraction, and risk-of-bias assessment, with screening the most mature [28]; reviews of AI-based tools report varying success and a consistent need for extensive human validation [29]. Existing tools address the downstream stages of review production; none addresses whether a review should be undertaken at all.

The system would combine 3 components. Semantic similarity modelling using dense vector embeddings would identify conceptually overlapping protocols and published reviews, going beyond keyword matching to capture PICO-level similarity, made tractable at registry scale by sentence-level embedding architectures [55]. PICO entity extraction using fine-tuned models would parse population, intervention, comparator, and outcome elements from abstracts and protocols, enabling structured rather than global overlap assessment, for which annotated corpora already exist [56]. Evidence gap quantification would map the proposed question against existing reviews, trial registries, and prior gap-mapping work to estimate whether the evidence base is saturated, partial, or absent [57,58]. Output would be graded rather than binary, ranging from a high-overlap warning where near-identical reviews exist to a no-adequate-evidence-base designation. Deployment would occur at PROSPERO preregistration screening, journal editorial triage, and funder proposal review.

Significant methodological work remains before responsible deployment. Current tools perform promisingly on curated benchmarks but have not been validated across the heterogeneous landscape of clinical research questions, where terminology variation and reporting differences degrade performance [29]. Performance is sensitive to the embedding model and training corpus, which are domain-specific and not interchangeable [55,56], and reproducibility is a further concern, since model behavior depends on training data, architecture, and preprocessing choices that are often incompletely reported [30]. Evidence-gap identification faces the same difficulty, with heterogeneous, unstandardized approaches and no agreed reporting conventions [57]. Domain-specific validation frameworks, including benchmark datasets reflecting clinical research diversity and reporting standards enabling independent verification, are therefore a prerequisite rather than a refinement.

Living systematic reviews (LSRs), continuously maintained and periodically updated syntheses, shift from discrete one-off reviews to sustained evidence resources, combining moderate feasibility, a moderate evidence base, low-to-moderate scalability, and low risk [45,59]. Where a living review is maintained, the case for additional de novo reviews is substantially weakened, and updating becomes the default response to new evidence rather than a justification for a new publication [46]. COVID-19 supports this: living guidelines incorporated new trial results within 14 days whereas reviews of the same drug published over the same period were frequently obsolete at publication [10]. Evidence maps complement this by characterizing available evidence across broad topics to guide prioritization and identify saturation [58,60]. The barrier is sustainability: LSRs need ongoing funding and institutional buy-in absent from a model awarding funding to discrete projects. The original articulation assumed a continual approach could be sustained with feasible resources [45], but this remains unproven at scale, and sustaining a living review on a saturated question merely perpetuates attention on it [21]. Scaling LSRs would require fundamental restructuring of synthesis funding and the coordination infrastructure the ecosystem lacks [24,53].

The most transformative but most challenging interventions operate at policy level: funder mandates requiring evidence needs assessments before funding new reviews, journal policies requiring explicit novelty justification with systematic scoping evidence, and promotion criteria rewarding synthesis quality and coordination over publication volume [1,24,34]. These combine low-to-moderate feasibility, a limited evidence base, high scalability, and significant risk. The fundamental tension is between coordination and the epistemic value of independent replication: multiple syntheses can serve a validation function when primary evidence is uncertain, and overlap frequently reflects different questions legitimately addressed to the same literature [25,26]. Policy interventions that suppress redundancy therefore risk suppressing the replication that makes evidence synthesis epistemically robust. The record is cautionary: the research-waste agenda has been articulated since 2009 [6,15], yet measured redundancy has not visibly declined [5]. Policy reform is therefore most likely to succeed when bundled with measurement and validation work rather than as a stand-alone mandate.

The highest-impact near-term intervention is the combination of enhanced preregistration scoping mandates and AI-assisted overlap detection because together they address the production and governance dimensions with moderate complexity and strong scalability and because demand for decision support is documented [20]. Preregistration mandates are feasible immediately, since registry infrastructure and reporting templates exist; AI-assisted detection is scalable in principle but requires validation first [29,30]. Living reviews and policy reform are longer-term structural changes requiring substantial institutional investment, their benefit conditional on stable funding and coordination the system does not supply [24,53]. Two considerations temper this ranking: no intervention has been evaluated in a controlled trial, so the ranking reflects strength of reasoning rather than demonstrated effect; and the interventions are not independent, since scoping mandates without detection are unenforceable at scale, while detection without a mandate produces no behavioral change.