Curbing the tide of "paper mills": A risk-based integrated governance for public health databases

Large-scale public health databases like the US National Health and Nutrition Examination Survey (NHANES), are increasingly being misused. Researchers generate large volumes of low-quality papers by mechanically combining variables and conducting shallow correlation analyses while using AI tools to bypass plagiarism checks. A recent methodology and cross-sectional study published in The BMJ revealed that nearly 10% of original cancer research papers, exceeding 260,000 articles, exhibited textual signatures consistent with retracted paper-mill outputs. Crucially, this study found that the prevalence of such “flagged” papers has grown exponentially in high-impact journals (top 10% by impact factor), highlighting that traditional peer review is struggling to curb the tide of industrialized fabrication.2 The Retraction Watch database has documented tens of thousands of retractions linked to systematic fraud. Recent analyses further reveal that paper mills are increasingly exploiting publicly accessible datasets, such as NHANES and FAERS, to generate large volumes of low-utility, single-association studies. Meanwhile, the United2Act initiative, a multi-stakeholder collaboration supported by COPE and STM, has highlighted the unintended consequences of open-data policies. While these policies have significantly advanced equity and transparency, they have also created opportunities for industrialized research misconduct. Consequently, the initiative calls for robust, coordinated governance involving database curators, publishers, and research integrity organizations.

In response, journals such as Frontiers and PLOS now often reject submissions relying on these databases unless the authors provide external experimental validation or institutional supplementary data. Simultaneously, databases such as jPOST are addressing these issues by developing data quality assessment indices and establishing specialized data journals to promote responsible use. The UK Biobank also issued a statement suspending data access until download restrictions are in place. A 2024 study offers a comprehensive overview of ongoing efforts at both the publisher and organizational levels to combat paper mills. These efforts include layered detection tools, such as network analysis and image forensics, as well as shared toolkits developed by independent organizations and practical checklists designed for editors and readers.

Despite these efforts, Dr. Richardson shows that the entities enabling scientific fraud at scale remain large and resilient and continue to grow rapidly. They achieve this growth primarily through tactics such as journal hopping and broker networks. Therefore, to combat paper mills, it is essential to establish a collaborative governance framework involving journals, databases, and other stakeholders. Dr. Sanderson described a major multi-stakeholder initiative involving funders, publishers, and research institutions, aimed at systematically addressing the commercial practices behind the production of fraudulent academic papers. He emphasized that countermeasures must shift from isolated efforts by individual journals to a coordinated, industry-wide response.

Building on these existing studies, we believe a risk-based automated screening and information-sharing system should be established, leveraging the technological feasibility demonstrated by recent AI models. The BMJ study mentioned above successfully utilized a fine-tuned BERT (bidirectional encoder representations from transformers) model to classify paper mill products with 0.91 precision using only titles and abstracts.2 Although this example demonstrates feasibility, any automated risk-scoring system is likely to be reverse-engineered quickly. Paper mills adapt rapidly by paraphrasing content, distributing queries across multiple accounts, or mimicking the patterns of legitimate submissions. Recent evaluations indicate that even the most advanced AI detectors remain vulnerable to hybrid adversarial edits. This vulnerability underscores the need for continuous retraining and sustained hybrid human-AI oversight. To address these challenges, we propose a more robust ensemble machine-learning framework (Figure 1). The approach integrates natural language processing of manuscript content with behavioral anomaly detection at the database level.