Forty-Five Editors, Thirty Percent of the Retractions

Forty-Five Names on Two Spreadsheets

Two spreadsheets did it. Not a whistleblower, not a leaked internal email, not a confession extracted under pressure. Two ordinary bibliographic exports: one listing which handling editor accepted which manuscript at a journal, the other listing which of those manuscripts were later retracted or drew a comment on the post-publication review site PubPeer. Merge them. Run a statistical test on every editor who has ever worked there. See who comes out anomalous.

At PLOS ONE, which is unusual among large journals in publicly naming the editor who handled each accepted paper, that test flags forty-five people [1]. Forty-five out of 18,329 editors who have accepted at least one article since the journal launched in 2006, a group so small it is 0.25% of everyone who has ever done the job. Between them they have handled 1.3% of the journal's 276,956 published articles. And they account for 30.2% of everything the journal has ever had to retract.

That is the headline of a paper by Reese Richardson, Spencer Hong, Jennifer Byrne, Thomas Stoeger, and Luís Amaral, published in the Proceedings of the National Academy of Sciences in August 2025, titled, with unusual bluntness for a PNAS paper, "The entities enabling scientific fraud at scale are large, resilient, and growing rapidly" [1]. One of the authors, Amaral, is himself a longtime PLOS ONE editor, which is part of why the paper reads less like an outside accusation and more like an institution auditing itself and not liking what it finds.

What follows is our attempt to actually understand the machinery behind that number, rather than just repeat it. How do you catch forty-five people doing this without anyone confessing? Why is a network of colluding editors so much harder to catch than one person faking a gel image? And once you have caught it, what are you supposed to do with your trust in the review process the rest of us rely on to filter what counts as science?

Defection, Not Villainy

The paper's own framing is worth taking seriously before the numbers, because it changes what question you should be asking. Richardson and colleagues borrow a term from game theory: defection [1]. Treat the scientific enterprise as a public-goods game, in which everyone benefits from a shared pool of trustworthy knowledge and everyone is supposed to contribute genuine work to it. A defector takes the benefits, career credit, funding, citations, without contributing genuine knowledge. Historically, defection was assumed to be a rare, individual failure: one person under career pressure who fabricates one dataset.

The paper's taxonomy has three moving parts, and they are not the same thing. Paper mills mass-produce fake or low-quality manuscripts for sale. Brokers are the middlemen who connect that supply to journals willing to publish it, sometimes by paying editors, sometimes by embedding themselves as editors. Predatory or hijacked journals are the venues, sometimes legitimate journals that a broker has captured outright, that skip the quality control a journal is supposed to provide. None of the three, alone, needs a network of forty-five people. What needs the network is doing all three at once, repeatedly, without getting caught, at a scale bigger than any single fabricator could manage alone.

That reframing matters for how you read every number that follows. This is not a story about forty-five unusually dishonest individuals. It is a story about what a public-goods game looks like once defection stops being a private, individual failure and starts being an organized, cooperative strategy.

How You Catch an Editor Doing This

Here is the actual statistical machinery, because it is more interesting than "they found some bad editors."

PLOS ONE has published 276,956 articles since 2006. Of those, 702 have been retracted and 2,241 have drawn a comment on PubPeer [1]. Neither number is evenly spread across the journal's history, and neither is evenly spread across its 18,329 editors. For every editor, for every year they were active, the authors counted how many of the manuscripts that editor accepted were eventually retracted, and separately how many eventually drew a PubPeer comment. Then they asked a precise statistical question: given the journal-wide, time-varying background rate of retraction and PubPeer commenting, is this editor's count higher than chance would produce? The tool is a Poisson binomial test, run editor by editor, with the false discovery rate held at 0.05 across all 18,329 simultaneous tests [1] — the same correction problem you'd face testing thousands of genes at once, applied here to thousands of editors at once.

They ran the test three separate ways, because they had three separate signals available: eventual retraction, eventual PubPeer comment, and unusually short turnaround time between submission and acceptance (30 days or fewer), which the literature treats as one hallmark of paper-mill-adjacent processing [1]. The three signals did not agree equally well. Retraction anomalies flagged 22 editors; PubPeer anomalies flagged 33; short-turnaround anomalies flagged a striking 88. But when the authors checked whether authors were also disproportionately routed to those flagged editors, the numbers inverted: 21 authors for the retraction-flagged group, 18 for the PubPeer-flagged group, and only 7 for the short-turnaround group. Short turnaround, in other words, catches a lot of editors but doesn't cleanly separate a coordinated network from ordinary journal variation — and a footnote from Amaral explains part of why: PLOS ONE itself actively encouraged fast turnaround for a period, a policy since deprioritized, so some editors were simply doing what the journal asked [1]. The authors' own data show their weakest signal being weak for a documented, non-nefarious reason. That is the kind of honesty a paper claiming to have caught fraud needs to have.

The Ones Who Send Work to Each Other

The 22, 33, and 88 above overlap, and they don't capture everyone. An editor can also be flagged indirectly: by having their own submissions disproportionately handled by other flagged editors. Combine every direct and indirect route and the total comes to 45 PLOS ONE editors [1] — the number in this piece's title.

Here is the club's own arithmetic on top of that. If retractions were distributed in proportion to how many articles a group of editors actually handled, 45 editors responsible for 1.3% of PLOS ONE's articles should be responsible for about 1.3% of its 702 retractions: roughly 9. They are instead responsible for 30.2%, or about 212. Divide 212 by 9 and you get a group producing retractions at more than 23 times the rate their share of the workload would predict. That ratio is ours, not the paper's; the paper reports the two percentages, we divided them.

What makes this a network finding rather than forty-five separate individual findings is what the editors did with each other. More than half, 25 of the 45, also personally authored articles that PLOS ONE went on to retract [1] — the same people show up on both sides of the review, as gatekeeper and as submitter. And when the authors mapped who sent submissions to whom, they found a densely connected cluster of editors serving between 2020 and 2023, based across four different countries, who sent most of their own submissions to each other rather than to the journal's other 18,284 editors [1]. More than half of the articles this cluster accepted were later retracted, and the retraction notices for that group read almost interchangeably:

A journal's own editorial staff, in other words, eventually recognized this as a series — a repeated pattern, not isolated incidents — even before the PNAS analysis formalized it statistically. The paper is careful to note one honest complication here, and we think it deserves repeating rather than smoothing over: editor assignment at PLOS ONE is nominally close to random, but authors are permitted to recommend an editor, so an author landing with a preferred editor "is not unlikely" on its own [1]. Some of what looks like editor-side self-dealing could in principle be author-side targeting of accommodating editors. The statistical signature doesn't fully distinguish "editors who favor each other" from "authors who know which editors to ask for," and the paper doesn't pretend it does.

Same Story at Other Journals, and in Conference Halls

PLOS ONE is unusual only in that it publishes editor names, which is what made this particular test possible there. Hindawi, the other publisher that discloses handling editors, offers a second, independent test of whether the pattern generalizes. Restricting to the ten Hindawi journals with the most retractions and PubPeer comments, the same statistical procedure flags 53 editors for anomalous retraction rates, 52 for anomalous PubPeer rates, and 96 for anomalous fast-tracking, with six, four, and seven of those, respectively, flagged at more than one journal simultaneously [1] — meaning some individuals were doing this across multiple Hindawi titles at once, not just one.

The same method extends to a venue with no human "editor" in the usual sense at all: IEEE conference proceedings. Treating each conference's organizing committee as equivalent to an editor, the authors examined 2,294,067 articles published across 19,969 IEEE conferences since 2003. Forty-five conferences held between 2009 and 2011 had more than 10% of their published articles eventually retracted, producing a visible spike in the retraction record around those years [1]. Applying the full three-signal test conference-wide flags 84 conferences for retraction anomalies, 158 for PubPeer anomalies, and 183 for anomalously frequent "tortured phrases," the garbled synonym-substitutions typical of paraphrasing software used to evade plagiarism detection. At least seven multi-year IEEE conference series were flagged for tortured-phrase anomalies in every single year they were held [1] — the same organizing apparatus, repeating the same behavior, year after year, for years.

A Second Fingerprint, Left in the Images

Editorial handling records are one kind of evidence. The paper builds an entirely separate line of evidence from something with no editor attached at all: duplicated images. Starting from PubPeer reports of image duplication, the authors constructed a network in which the nodes are articles and the edges are instances of two articles sharing a duplicated image. That network holds 2,213 articles connected by 4,188 such duplicate-image observations, splitting into 20 separate connected components, the largest of which contains 622 articles on its own [1].

Because some of those components are large, the authors used a stochastic block modeling approach, a statistical method for finding densely connected sub-groups (modules) inside a bigger network, to look for internal structure [1]. Two things fell out of that analysis that matter more than the network's raw size. First, the articles inside a given module are not randomly distributed across publication years or across publishers; a Shannon-entropy test rejects that null hypothesis for essentially every module with enough data to test it [1], meaning a batch of image-linked articles tends to land in the same journal in a tight time window — exactly what you'd expect if a broker were placing a shipment of manuscripts rather than individual authors independently reusing their own old figures. Second, despite every article in this network showing hard visual evidence that its data could not be what it claims, only 34.1% has actually been retracted [1]. Two-thirds of a network built entirely out of provably duplicated images is still sitting in the scientific record with no correction attached to it.

Why Not Just Check Who Cites Whom?

Here is a natural guess about how you'd catch a cartel of cooperating editors, and it is worth confronting directly because it is not what this paper did. If a group of editors and authors were working together, wouldn't they also be citing each other, inflating each other's citation counts and journal impact factors the way known citation cartels do elsewhere in the literature? Coordinated editorial action toward exactly that goal, boosting a journal's impact factor through engineered citation patterns, is a real, separately documented phenomenon; a 2022 study in Research Policy examined how institutions and journals game citation-based rankings this way [2], and a 2024 Scientific Reports paper specifically mapped fabricated co-authorship networks used to sell authorship slots on manuscripts [3].

But that is not the fingerprint Richardson and colleagues actually pulled up. Nowhere in their PLOS ONE, Hindawi, or IEEE analyses does a citation network appear. The signal that caught this particular network was who accepted whose manuscript and who reused whose image, not who cited whose paper. That distinction is not a technicality; it changes what an editorial fraud ring actually has to do to stay invisible. A citation cartel has to cite. It leaves tracks in exactly the dataset bibliometricians already scrutinize for this kind of gaming. A network that fast-tracks and rubber-stamps each other's submissions, by contrast, doesn't need its members' papers to reference each other at all — the papers can be about completely unrelated topics, in completely different subfields, and never once cross-cite. Which means an audit that only ever looked at citation graphs, however sophisticated, would have sailed straight past this one. The editorial layer and the citation layer are different crime scenes, and this paper worked the one that citation analysis doesn't reach.

The Business That Follows the Deindexing List Around

The paper's clearest single piece of evidence for an organized broker, as opposed to a loose informal network, is a case study of an entity it names outright: ARDA, a business whose website as of June 2024 advertised "Conferences and Meetings," "Journal Publication," and "Thesis/Article Writing," and claimed involvement in "4,565+ publications" [1].

Since January 2018, ARDA has listed 188 unique journals on its website as guaranteed publication venues, and the list is not static. It grew from 14 journals in January 2018 to 86 by March 2024 — the club's own arithmetic on those two endpoints works out to roughly 34% compound annual growth in portfolio size, sustained over six years. Of the full 188-journal list, 106 (56.4%) were indexed by Scopus, 51 (27.1%) by Web of Science, 29 (15.5%) by both, and just 2 (1.1%) by MEDLINE; 17 of the 188 (9.0%) are suspected "hijacked journals," formerly legitimate titles a paper mill has taken complete editorial control of [1].

What makes ARDA read as an active operator rather than a passive list-keeper is how the portfolio moves in response to enforcement. A journal ARDA had listed in 2020 and that was Scopus-indexed at the time had a 33.3% chance (13 of 39) of being deindexed later, versus a 0.5% baseline chance (147 of 27,197) for a Scopus-indexed journal actively publishing in 2020 generally — a difference the paper confirms is not chance with a two-tailed test (P < 0.0001) [1]. Dividing those two percentages, as the club did, puts an ARDA-listed journal at roughly 66 times the deindexing risk of an ordinary journal. And when a batch of journals was deindexed by Scopus in 2020 or 2021, ARDA removed them from its site by May 2021 and replaced them with new titles [1], exactly the behavior you'd expect from an operator actively managing a portfolio around enforcement action rather than one that simply happened to pick unlucky journals. For a set of five ARDA-listed journals the authors inspected in detail, between 34.0% and 98.7% of published articles fell entirely outside the journal's own stated scope — the hazelnut-roasting paper in an HIV/AIDS journal, the malware-detection paper in a special-education journal [1].

Some Fields Are Simply Better Targets

A separate section of the paper asks whether fraud is spread evenly across science, and answers no, using a clean internal control. Restricting to six closely related, similarly sized subfields within RNA biology, all with comparable growth trajectories since roughly 2000, the authors tracked annual publication counts alongside annual rates of errata and retractions from 2002 to 2022 [1]. Errata, corrections to honest mistakes, sit in a narrow band of 1.5% to 2.5% across every one of the six subfields; the literature in each is being revisited for ordinary reasons at roughly the same rate [1].

Retraction rates do not share that consistency. CRISPR-Cas9 sits near 0.1%. Studies of transfer RNA and cancer peak around 1%. Circular RNA studies peak near 2.5%. Both microRNA-and-cancer studies and long non-coding RNA studies peak near 4%, and for individual journals publishing in those two subfields specifically, the retraction rate exceeds 10% [1]. Because the errata rate, the honest-mistake baseline, stays flat while the retraction rate swings by a factor of 40 across subfields studied with the same tools over the same period, the authors argue this gap cannot be a good-faith artifact of a field being newer or less mature. It reads as targeting: paper mills preferentially select certain subfields, likely because reviewers there are less equipped to spot fabricated results, or because demand for cheap, fast publication is higher in fields under acute career pressure [1].

The Arithmetic of Outrunning Everyone Chasing It

The paper's closing empirical argument is about speed, and it is the number that should worry a reader most. Comparing four annually-growing quantities over recent decades, the authors report doubling times of 15.0 years for total global scientific output, 3.6 years for PubPeer-commented articles, 3.3 years for retracted articles, and just 1.5 years for suspected paper-mill products [1].

Doubling times are awkward to compare by eye, so here is the club's own conversion. A doubling time T corresponds to a continuous annual growth rate r = ln(2) / T. Run that arithmetic on the paper's four numbers and the annual growth rates come out to roughly 4.6% for total science, 19.3% for PubPeer comments, 21.0% for retractions, and 46.2% for suspected paper-mill products. Divide the fastest by the slowest and the fraud economy, as the paper's own trend lines describe it, is growing at almost exactly ten times the rate of legitimate science itself.

Doubling times: Richardson et al. 2025 [1]. Growth-rate column and the 10× ratio: club arithmetic, r = ln(2)/T.

The consequence the paper draws out is that the mechanisms meant to catch up are themselves outmatched. Web of Science and Scopus each deindex on the order of a hundred journals per year, a number the authors note is tenfold smaller than the number of journals currently publishing suspected paper-mill products [1]. Of the 29,956 suspected paper-mill products in their corpus that have a matching OpenAlex record, only 8,589, or 28.7%, have actually been retracted; extrapolating current trends, the authors project that only around 25% of today's suspected paper-mill products will ever be retracted, and only around 10% will ever end up in a deindexed journal [1]. Put plainly: on the paper's own projection, roughly three out of every four papers already identifiable as probable paper-mill products today will sit in the permanent scientific record, uncorrected, indefinitely.

Why This Beats the Immune System Built for Lone Fabricators

It is worth being precise about why a coordinated network like this is a fundamentally harder catch than one person faking a Western blot, because the two failure modes break the same accountability system in opposite ways.

A lone fabricator concentrates risk. One person, one paper, one implausible dataset, one image reused where it shouldn't be. The signal is small in scope but large in magnitude, which is exactly what post-publication review sites like PubPeer are built to catch: a sharp-eyed reader notices that one figure looks wrong. A coordinated network does the opposite. It distributes risk unevenly, by design or by accident. Of the 45 flagged PLOS ONE editors, only 22 carried an individually anomalous retraction rate on their own record — a needle so fine, out of 18,329 editors, that finding it required a purpose-built statistical test run simultaneously on everyone who has ever done the job, with a false-discovery correction to stop the multiple-testing problem itself from manufacturing false positives. The other 23 needed no anomalous personal record at all. They were caught only because they routed submissions to editors who were already flagged, a second, independent signal that has nothing to do with any one person's own numbers looking alarming.

That is the actual asymmetry. Catching one bad paper takes a reader with good instincts. Catching this took a research team, a full retraction database, PubPeer's comment archive, and a statistical method sophisticated enough to separate a real coordinated signal from eighteen thousand editors' worth of ordinary noise, and even then, the network the authors describe operated for at least three years, 2020 to 2023, before this analysis surfaced it. A network built around distributing suspicion thinly is, almost by construction, a network built to survive the detection tools designed around a lone, concentrated defector.

What This Should, and Should Not, Change About Trusting Peer Review

It would be easy to walk away from these numbers with a flattened takeaway: peer review is broken, don't trust journals. The paper itself resists that reading, and we think it's right to. Forty-five editors is 0.25% of everyone who has done the job at PLOS ONE. The other 99.75% show no statistical anomaly at all. The finding is not that editors, as a class, cannot be trusted. It is that the accountability system built around the assumption that fraud is rare and individual has no real capacity to catch fraud that is organized and cooperative, and that gap is being actively exploited.

The paper's discussion section makes a specific, structural argument about why, and it is worth quoting rather than paraphrasing away. Journal editors, the authors note, "have been offered substantial payments for rapid publication of selected manuscripts," and separately, "coordinated editorial action has been implicated in efforts to increase the impact factor of journals" [1], citing among other work the impact-factor gaming literature discussed above [2]. The people currently doing the work of finding this, meanwhile, are described by the authors as "courageous but isolated individuals," some of whom have "been accused of vigilantism and dismissed," others "threatened with legal action" [1]. The authors' own recommendation is structural, not moralistic: separate the tasks of detection, investigation, and sanctioning, and remove each from the hands of parties who have a conflict of interest in the outcome, because right now, a university investigating its own star researcher, or a publisher investigating its own flagship journal, is being asked to police itself [1].

There's a coda in the discussion that we think is the most quietly alarming sentence in the paper, and it has nothing to do with any individual editor. The authors point out that both symbolic "machine scientists" and large language models are increasingly used to summarize and synthesize the scientific literature for researchers and the public, and that none of those tools can currently distinguish a real paper from a fraudulent one [1]. Every fraudulent paper this network produced, and every one still uncaught, becomes an input to systems built on the premise that the literature they're reading is trustworthy. The fraction of the corpus that isn't is, by the paper's own numbers, the fastest-growing fraction there is.

What the Authors Themselves Flag as Limits

In fairness to the paper, its own limitations section is unusually direct, and worth repeating rather than skating past. The analysis can only find fraud that has already left a trace, retraction, a PubPeer comment, a duplicated image, in the corpora the authors could access, and the consensus among experts they cite is that the large majority of paper-mill products have not yet been detected at all [1]. Several of the case studies, ARDA, the RNA biology subfields, focus on particular businesses or disciplines that may not generalize to fields the authors did not examine. And changes over time in how much attention different fields or journals receive from watchdogs could themselves produce trends that look like growth in fraud without being growth in fraud [1] — a possibility the authors say they cannot fully rule out, only argue against on balance.

We'll end where the paper's authors end their own discussion: not with a verdict that the system has failed, but with an argument that detecting fraud at the scale this paper describes cannot keep being the work of "isolated volunteers." It needs resources, both human and technological, that are actually commensurate with a fraud economy currently growing about ten times faster than the science it's hiding inside of.

References

1. 1.

   Richardson, R. A. K., Hong, S. S., Byrne, J. A., Stoeger, T. & Amaral, L. A. N. (2025). The entities enabling scientific fraud at scale are large, resilient, and growing rapidly. Proceedings of the National Academy of Sciences 122(32), e2420092122. doi:10.1073/pnas.2420092122

2. 2.

   Siler, K. & Larivière, V. (2022). Who games metrics and rankings? Institutional niches and journal impact factor inflation. Research Policy 51(10), 104608. doi:10.1016/j.respol.2022.104608

3. 3.

   Porter, S. J. & McIntosh, L. D. (2024). Identifying fabricated networks within authorship-for-sale enterprises. Scientific Reports 14, 29569. doi:10.1038/s41598-024-71230-8

4. 4.

   The PubPeer Foundation. PubPeer — post-publication peer review. https://pubpeer.com. Comment archive cited as a primary data source in [1].

5. 5.

   The Center for Scientific Integrity. Retraction Watch Database. https://retractiondatabase.org. Retraction corpus (47,387 records as accessed) cited as a primary data source in [1].

6. 6.

   Committee on Publication Ethics (COPE) & International Association of Scientific, Technical and Medical Publishers (STM) (2022). Paper mills: research report from COPE and STM.