A familiar database name is not a quality label: Correspondence argues that the risk in public database research lies in how science is produced

As a scientometric analysis estimates more than 11,500 excess publications from nine heavily reused health datasets in a single year, a correspondence in Medicine Discovery argues that the databases themselves are not the problem.

Public biomedical databases have become a focus of the research-integrity debate. In recent years, editors, publishers, and major scientific media have drawn attention to a rapid rise in papers built from broadly accessible datasets, with resources such as the National Health and Nutrition Examination Survey (NHANES), UK Biobank, the FDA Adverse Event Reporting System (FAERS), the Global Burden of Disease (GBD) Study, and FinnGen among those most frequently discussed. A correspondence published in Medicine Discovery argues that this debate risks conflating two different things: scientific infrastructure built to widen access, and the production model that exploits it.

The correspondence titled "Old data, new problem: rethinking public database research in the era of generative AI" states its central claim plainly at the outset. As the authors write, "Our central argument is that the quality of public database research rests on the study process, while the database name provides little information about scientific quality."

A rapid change in publication volume

A recent scientometric analysis by Spick, Onoja, and colleagues examined 36 health datasets and identified nine whose publication patterns were consistent with rapid, formulaic expansion. The study searched OpenAlex on 3 January 2026 for article records dated from 2014 through 2025, using dataset terms in titles or abstracts, and applied 2014 through 2022 as a forecast baseline; a 2025 count above the 95% confidence interval was classified as unusual growth. The nine datasets produced 23,005 indexed publications in 2025, and an autoregressive integrated moving average forecast estimated an excess of 11,577 publications above the prior trend, compared with 7,655 publications in 2022.

The correspondence treats those figures as a description of publication volume rather than a verdict on any individual paper. It points to a parallel line of evidence from Suchak and colleagues, who documented a sharp increase in formulaic NHANES papers and identified recurrent problems including single-exposure analyses, biologically implausible claims, selective use of survey cycles, and weak control of false discovery. The authors characterise these as failures of question formulation, study design, statistical reasoning, and publication practice rather than as defects of the data source. As they write, "A dataset can enable excellent science and weak science at the same time."

Databases are not the problem

The correspondence argues that the database itself carries little information about quality, and that editorial screening can inadvertently turn a data source into a proxy for it when that source is visible in the title, abstract, or methods. NHANES grew out of a long-standing United States national health examination programme and became a continuous survey in 1999. Its design combines interview data with examinations and laboratory measurements under a complex probability sample, which supports nationally informative health estimates, and current guidance from the National Center for Health Statistics emphasises cycle-specific analytical decisions as well as survey weighting and variance estimation. Researchers have worked with these data for decades, and the recent surge in publications emerged long after the survey architecture and its public documentation were established. In the authors' reading, the timing points towards a change in the research production environment rather than a change in the resource.

UK Biobank offers what the correspondence describes as the same lesson at a larger scale. Recruitment began well before the current generative AI era, and the resource was developed to support wide-ranging research on the causes of disease through deep phenotyping and long-term follow-up; its scientific value comes from scale, linkage, repeat measurements, imaging, genetics, biomarkers, and prospective outcomes. A high publication count is therefore an expected consequence of a resource designed for broad reuse, and volume alone carries little information about methodological quality. A strong paper, the authors note, may ask a difficult question using a familiar cohort, while a weak paper may apply a fashionable method to a widely studied question using a novel dataset.

What generative AI changed

Generative AI, the correspondence argues, has changed the economics of secondary data research. Large language models can support many steps of a secondary data project, from code generation to manuscript drafting. A 2023 evaluation showed that ChatGPT could produce plausible scientific abstracts that human reviewers sometimes had difficulty distinguishing from authentic abstracts; the authors note that this supports the capability of language generation, while broader effects on analytical quality and publication speed require separate evidence. A separate analysis of publishing found that papers published after 2022 became longer and cited more references. Researcher-directed AI assistance can support bounded tasks such as coding or language editing, whereas more autonomous workflows can extend AI involvement into question formulation and analytical planning.

The correspondence distinguishes assistance from autonomy in explicit terms. As the authors write, "Legitimate AI assistance is compatible with responsible research when researchers retain scientific control and report its use transparently." Publication-driven workflows, they add, can substantially reduce the time required for repetitive tasks, which creates a plausible route to rapid formulaic production even as the scale of that practice remains an empirical question.

That route becomes more efficient when it meets a standardised public dataset. A resource such as NHANES contains hundreds of variables, repeated survey cycles, familiar outcomes, and extensive documentation. Once a reusable analysis pipeline has been built, exposure A can be replaced by exposure B, outcome X by outcome Y, and subgroup M by subgroup N; automated code can cycle through models, language models can produce narratives, and a manuscript can acquire the surface features of a conventional epidemiological paper. When the space of possible questions is large, statistical significance becomes easy to search for, producing what the authors describe as a plausible risk of automated hypothesis fishing. Their concern is directed at the research workflow surrounding open data rather than at open data itself.

Consequences for training and research culture

The correspondence argues that the shift has direct consequences for young researchers. Public datasets once offered an affordable way to learn epidemiology, biostatistics, reproducible coding, and scientific writing, and the authors note that they still do; generative AI is also reshaping learning environments, with earlier work highlighting both its educational potential and the need for appropriate supervision to guide its use. The risk, in their view, lies in what the current publication culture rewards. When success is measured by the number of manuscripts generated from a dataset, trainees learn to optimise throughput; when supervisors and journals reward clear questions, defensible methods, replication, and transparent analysis, the same datasets become excellent training grounds. As the authors write, "We consider that the decisive variable is research culture."

Three actions for journals

The correspondence proposes three practical actions that editorial policy can take. First, journals can request a short dataset-specific methods note for studies based on heavily reused resources. In NHANES studies, the note can explain cycle selection and the approach to weighting and variance estimation; for reused cohorts, authors can add a brief statement describing overlap with earlier reports. The authors argue that such requirements can remain proportional to study scope, which preserves access for smaller teams and supports legitimate replication.

Second, journals can strengthen transparency around analysis and AI use. Public data studies can provide analysis code and enough metadata to reproduce variable construction when sharing is feasible. Preregistration is described as especially useful for confirmatory questions, while exploratory analyses can be labelled as exploratory and interpreted as hypothesis generating. Authors can also report whether generative AI contributed to question formulation or analytical planning, consistent with guidance from the International Committee of Medical Journal Editors on the disclosure of AI use and continued human responsibility for submitted work, and with guidance from the Committee on Publication Ethics on transparent reporting of AI use.

Third, the correspondence argues that editorial screening can focus on study conduct rather than on provenance. Reviewers can examine whether the analytic sample follows from the question being asked and whether the data structure has been handled correctly, and can assess the treatment of multiple testing and the role of sensitivity analyses; similarity tools can help identify highly repetitive submissions. Borderline cases, the authors suggest, can receive independent methodological review or an editorial appeal. Their stated conclusion is that editorial decisions should rest on scientific assessment instead of the database name.

What the correspondence does not claim

The correspondence is a commentary based on published literature rather than a report of new experimental data, and the authors present no new empirical analyses of their own. The growth figures it cites come from a scientometric forecasting analysis of indexed publication records and describe publication volume rather than the scientific quality of individual articles; an excess above a forecast trend is not, by itself, evidence that any particular paper is unreliable. Likewise, the observation that papers became longer and cited more references after 2022 describes a pattern in the literature and does not establish that AI tools caused a change in research quality. The governance measures the authors propose are suggestions for editorial policy rather than interventions that have been tested for effect, and the dataset examples they discuss are illustrative. The authors emphasise that the same public resources continue to support careful epidemiology, surveillance, methods development, replication, and policy-relevant research, and that a familiar database name should not be read as evidence of low quality any more than the visible use of AI should be read as evidence of misconduct. Any editorial response, they argue, should remain proportionate to study scope.

Looking ahead

As the authors write, "Generative AI has lowered technical and linguistic barriers to manuscript production, which gives human scientific judgment greater importance." NHANES, UK Biobank, FAERS, GBD, FinnGen, and similar resources remain valuable scientific infrastructure, and study quality arises from the reasoning and methodological care applied to each analysis. The correspondence concludes that editorial policy can respond through dataset-specific documentation and transparent reporting, while scientific review should remain centred on each individual study.

Article information:

Old data, new problem: rethinking public database research in the era of generative AI.

Medicine Discovery, 2026.

Original article link: https://www.sciencedirect.com/science/article/pii/S3117928226000020

DOI: 10.1016/j.medd.2026.100002

About the journal

Medicine Discovery (ISSN: 3117-9282) is a peer-reviewed, international, open-access journal published by Zhujiang Hospital of Southern Medical University in collaboration with KeAi Publishing and distributed on the ScienceDirect platform. The journal is dedicated to publishing high-quality translational medical research that bridges mechanistic discovery and clinical implementation, and covers five sections spanning mechanisms and systems medicine, translational and clinical research, digital medicine and data-driven healthcare, therapeutic innovation and biomedical engineering, and methodology, standards, and implementation science.

Website: https://www.sciencedirect.com/journal/medicine-discovery

Submit: https://www.editorialmanager.com/medd/default2.aspx