Thousands of near-identical health studies are appearing. Most may not be from paper mills

Anyone who follows medical literature has noticed the pattern: a steady stream of papers reporting the burden, trend and projection of some disease in some population, drawn from the same public dataset, structured identically, differing mainly in which disease and which population.

A study published on 20 August in the Journal of Clinical Epidemiology set out to determine where those papers come from. Its answer is more uncomfortable than the obvious one.

How the investigation started

The authors selected the Global Burden of Disease Study database after noticing something about their own work: a paper of theirs on Bayesian age-period-cohort models was experiencing a rapid surge in geographically clustered citations from GBD papers with Chinese affiliations [s1].

The question they framed is a genuine fork. Papers of this kind might originate from paper mills — commercial entities that sell authorships on mass-produced papers — or from the uncoordinated action of individual researchers, facilitated by AI tools and templated workflows [s1]. From the outside, the output looks similar. The distinction matters enormously for what should be done about it.

What they checked

The investigation combined several approaches. The authors collected bibliometric and article-level metadata from GBD papers looking for indicators of mass production; assessed 713 full-text articles for the R software version reported, whether code and data were available, and whether generative AI use was declared; qualitatively screened the figures of 180 articles for graphical similarities; and conducted an exploratory scoping investigation of online platforms, including social media sites and vendor websites, dedicated to do-it-yourself workflows for secondary analysis of public health data [s1].

The finding

They could not rule out paper mill involvement, and say so [s1]. But the evidence they assembled points substantially the other way.

A wide variety of R versions was listed across 477 articles [s1] — a pattern that argues against centralised production, since a mill running an industrial pipeline would be expected to leave a narrower software fingerprint.

The figures told a similar story. There were broad graphical similarities in figure styles, consistent with shared visualisation tools, but substantial variation in ancillary details, suggesting independent authors finalising their own figures [s1].

And the scoping investigation found the mechanism: an online ecosystem of proprietary tools and services specialising in streamlined do-it-yourself workflows for conducting, writing and publishing secondary analyses of public health data [s1].

The authors' conclusion is that the geographically clustered increase in GBD publications is at least partially driven by independent authors [s1] — people using purchased software and templates to produce their own papers, not people buying a byline on someone else's.

Why "not a paper mill" is not reassuring

Organised fraud is, in a sense, the easier problem. It has perpetrators, a business model, and detectable fingerprints. Journals can be alerted, papers retracted, and the specific pipeline shut down.

A commercial ecosystem selling legitimate-looking analytical workflows to individual researchers has none of those handles. Nothing in it is necessarily fraudulent. The data are public, the software is real, the analysis may be technically executed as described. What is being sold is speed and a template — and the result is a large volume of low-information papers that are individually defensible and collectively corrosive to the literature they enter.

The authors' recommendations follow that diagnosis rather than a fraud diagnosis. They argue that effort should go toward evaluating the quality of the identified workflows; that integrity stakeholders should monitor these online platforms; that code sharing should be mandated for data-driven secondary analyses; and that paywalls and proprietary software licences hinder transparency and reusability, which is why free and open-source software matters for trustworthy and reproducible research [s1].

Mandatory code sharing is the most concrete of these. A templated analysis is trivial to spot when the code is attached and trivial to conceal when it is not.

The reason it is worth caring about

Two pieces of reporting from earlier in the year explain why polluted literature is not a self-correcting problem.

Nature reported in July that paper-mill cancer studies receive double the number of citations as genuine papers [s3]. Whatever the mechanism, papers of questionable provenance are not sitting inert at the edge of the literature — they are being cited, which means they are feeding into reviews, meta-analyses and the evidence syntheses that clinical guidance is built from.

Science reported in August that hundreds of papers advertised for sale were subsequently published, with more than one-fifth of papers at some IEEE conferences appearing linked to authorship for sale [s2]. That concerns engineering venues rather than medical ones, but it establishes that the advertised-then-published pipeline is real and measurable where anyone bothers to look.

Detection efforts are underway — The BMJ covered the use of AI to identify paper-mill papers in March [s4] — though a detector trained on the signatures of organised mills is aimed at exactly the category this new investigation suggests may not dominate the GBD literature.

A caution about the geography

The clustering the authors observed is a real finding and they report it plainly [s1]. It is also the part most easily misread. Their conclusion is not that a national research community is committing fraud; it is close to the opposite — that individual researchers appear to be using commercially available tools independently [s1]. The vendors identified operate in a market, and markets go where demand and incentive structures send them.

What to watch

The practical test is whether journals adopt the code-sharing requirement the authors recommend for secondary analyses of public datasets [s1]. It is a low-cost policy with an obvious mechanism: it does not detect bad papers, it makes templated ones legible.

The harder question is what the GBD database's custodians and the wider open-data movement do about it. Public health data are made public for good reasons. This study is an early, careful account of what happens when the barrier to producing a publishable analysis from them falls close to zero.