AI-generated citation distortion: The trust crisis in biomedical publishing is spilling over into a global scientific governance issue

AI-Generated Citation Distortion: The Trust Crisis in Biomedicine Publishing Is Spilling Over Into Global Research Governance

An audit that appears to belong to the internal workings of academic publishing is revealing a broader global development issue: as generative AI enters the paper-writing process, the “low-cost expansion” of knowledge production and the “high-risk leakage” of evidence quality begin to occur at the same time. According to a correspondence in *The Lancet* and related reports, a research team audited about 2.5 million biomedical papers in the PubMed Central open-access database, covering the period from January 1, 2023 to February 18, 2026. In the end, they identified 4,046 fabricated citations among 97.1 million verified references, involving 2,810 papers.

This is not a marginal issue for the publishing world alone. It concerns whether medical evidence is trustworthy, whether systematic reviews are robust, whether clinical guidelines are reliable, and how research institutions in the Global South use knowledge systems amplified by digital platforms under constrained resources. More importantly, it reminds policymakers that AI brings not only productivity gains, but also reshapes the underlying logic of research integrity, knowledge verification, and international collaboration.

From “Writing Tool” to “Source of Evidence Pollution”

Generative AI was initially seen by many researchers as an assistive tool for improving efficiency: helping polish language, organize literature, and generate outlines. However, this audit shows that when AI is embedded in academic writing workflows without constraints, citation distortion can rapidly evolve from isolated mistakes into a large-scale risk.

The research team used an automated verification process, cross-checking databases such as PubMed, Crossref, OpenAlex, and Google Scholar, and used large language models to assist in screening suspicious entries. The method itself is significant in practice: at the scale of millions of papers, traditional manual verification is nearly impossible, and combining automation with human review has become a practical pathway for large-scale research governance.

But this is also where the problem becomes clear: once citation-checking mechanisms themselves lag behind, generative AI may inject content that merely “looks like evidence” into the knowledge system at scale. For the medical field, this distortion is especially dangerous because clinical guidelines, drug evaluation, and public health research all depend on a traceable and verifiable literature base. Incorrect citations are not just formatting flaws; they may affect treatment recommendations, policy judgments, and resource allocation.

Why This Is Not Only a Research Integrity Issue, but Also a Global Development Issue

From a global development perspective, research integrity is not an ivory-tower matter, but part of public capacity. For high-income countries, incorrect citations mainly mean damage to academic reputation; for low- and middle-income countries, they can also mean wasted limited research funding, contaminated policy evidence, and health systems absorbing unreliable evidence.

The World Bank, UNDP, and multilateral development institutions have repeatedly emphasized in recent years that “institutional capacity” and “data capacity” are important determinants of development outcomes.The World Bank, UNDP, and multilateral development institutions have in recent years repeatedly emphasized that “institutional capacity” and “data capacity” are important determinants of development outcomes. Today, this capacity is no longer confined to fiscal management or statistical systems; it also includes the knowledge production system itself: whether papers are verifiable, whether research is reproducible, whether citations are traceable, and whether databases are auditable.

This is highly relevant to the “governance” dimension of ESG. In the past, ESG was used more for corporate disclosure, supply chain management, and capital market assessment; now, governance issues surrounding AI-generated content are also beginning to fall within the responsibility of research institutions, publishers, and platform companies. If academic publishers cannot establish stronger reference verification mechanisms, they are effectively leaving a systemic governance gap in the knowledge supply chain.

The rise in fabricated citations reflects structural pressure in the research ecosystem

Audits show that the incidence of fabricated citations rose from about 4 per 10,000 papers in 2023 to about 57 per 10,000 papers in early 2026, an increase of more than 12 times. Although the proportion for any single paper may seem low, its cumulative effect in a million-paper environment and a multi-layer knowledge translation chain cannot be ignored.

This rise is not merely a technical issue; it is also related to the incentive structure of research. Globally, the number of papers, journal impact factors, grant success rates, and institutional rankings continue to strongly drive a “fast output” model. For some researchers, AI writing tools provide an incentive to cut costs and increase output; for paper mills, they may become an accelerator for mass-producing “academic appearances.”

The problem is that when publication pressure and technological convenience rise at the same time, while the verification capabilities of peer review, editing, and indexing systems do not upgrade accordingly, the research ecosystem develops a classic “quality externality.” That is why this phenomenon deserves to be understood within a global governance framework: it reflects not a single-point failure, but a mismatch among incentive mechanisms, platform technology, and institutional regulation.

What it means for developing countries

For universities, hospitals, and research institutions in many developing countries, the dual effects brought by generative AI are even more complex. On the one hand, it lowers language barriers and helps non-English-speaking researchers enter the international publication system more quickly; on the other hand, it also increases dependence on external databases, commercial tools, and opaque models.

If reference verification capacity is insufficient, local scholars are more likely to become trapped in “visibility competition” in the rush to publish, while neglecting evidence quality. More concretely, health systems, agricultural extension, climate adaptation, and education reform in developing countries often rely on international literature and external knowledge networks. Once these networks are contaminated with large numbers of fabricated or untraceable citations, decision-making costs will quietly rise.

This is also a new manifestation of the digital divide: in the past, the divide was reflected in network coverage and device access; now, it is also reflected in who can use high-quality verification tools, who can afford data auditing, and who can participate in rule-making.

Why publishers, indexing platforms, and international organizations must all intervene“The Lancet” correspondence and related institutions have suggested that publishers conduct citation checks at the submission stage, that indexing services add metadata flags to suspicious references, and that dedicated categories be created to track fabricated citations. These proposals are not exaggerated; rather, they represent the minimum governance requirements for academic infrastructure in the AI era.

The reason is straightforward: academic publishing is no longer an internal process within a single institution, but a global knowledge network spanning platforms, databases, and languages. Once systems such as PubMed, Crossref, OpenAlex, and Google Scholar become de facto gateways to knowledge, they assume a role akin to public infrastructure. Public infrastructure cannot pursue coverage alone; it must also pursue verifiability.

International organizations can also play a role here. Just as the climate field needs an MRV (monitoring, reporting, verification) system, research governance also needs a similar verification framework. In the future, cross-institutional standards may gradually take shape around transparent disclosure of AI-assisted writing, citation authenticity labeling, identification of paper mills, and database-linked early warning mechanisms. This will be a new frontier for international cooperation.

“Governance” from an ESG perspective is being redefined

In ESG discourse, academic publishing and research institutions are usually not regarded as traditional ESG assets. But if knowledge production is viewed as social infrastructure, it becomes clear that its governance quality directly affects healthcare, education, public policy, and investment decisions.

For investors, “governance risk” in the AI era no longer includes only board independence or financial disclosure; it also includes data sources, model usage, and content verification. For universities and research institutions, true competitiveness is not just the number of papers, but the integrity of the evidence chain. For publishers and database platforms, future core value will also lie not merely in scale of inclusion, but in the ability to manage trustworthiness.

This means that the ESG evaluation system itself may also evolve: as AI becomes involved in content production, governance indicators will need to cover algorithmic transparency, source traceability, error-correction mechanisms, and platform accountability. In other words, the governance dimension will expand from “internal corporate compliance” to “knowledge ecosystem compliance.”

The key to the future is not banning AI, but rebuilding verification mechanisms

The answer to these kinds of problems is not simply to return to the “pre-AI era.” Generative AI does have value in language support, knowledge retrieval, and research assistance, especially for non-English research communities. The real challenge is how to rebuild evidence verification and accountability tracing mechanisms while improving efficiency.

In the next few years, what is worth watching is not whether AI enters research, but three things:

1. Whether publishers move reference verification to the submission stage; 2. Whether indexing platforms establish unified labels for fabricated citations; 3. Whether research institutions incorporate AI-use disclosure and citation audits into research integrity systems.

If these mechanisms cannot be established, AI will continue to drive growth in research output, but the credibility of the knowledge system will be continually diluted. For medicine and public health, that is not an acceptable cost.## Conclusion: Knowledge Credibility Is Becoming a New Capacity for Development

Global development research has traditionally focused more on funding, infrastructure, education, and health; now it must bring “knowledge credibility” into the same framework. In an era of high digitalization, modeling, and automation, policy, healthcare, education, and investment increasingly depend on verifiable information flows.

This audit of biomedical papers shows that one of the scarcest resources in the AI era may not be content-generation capability, but content-verification capability. For countries in the Global South, this is especially important: if reliable knowledge infrastructure cannot be established, digitalization will not necessarily lead to more equitable development; instead, it may amplify existing inequalities.

Over the long term, what truly determines the competitiveness of a research system is not how fast it can write, but whether the evidence can be trusted.