Paper Mills and Self-Correction

Frontier If a blinded forensic audit were run on a random sample of several thousand papers drawn from one defined field and year — which nobody has done, and which section 7 specifies — the prize is one number with stated bounds in place of a fourteen-fold disagreement. The published estimates of how much of the literature is paper-mill output run from 2% of submissions, in the 2022 COPE and STM industry report, to 28.8% of a 2020 biomedical sample in the Sabel group’s preprint, with a machine screen putting 9.87% of 2.6 million cancer papers in the suspect class. Those are not three measurements of one quantity. They are three different quantities produced by three different instruments, none of which has been calibrated against a known answer, because no known answer exists.

Established The magnitude at stake is the gap between what enters and what is removed. Ten research-integrity specialists writing in Nature in January 2025 put at least 400,000 papers published between 2000 and 2022 in the paper-mill-hallmark class against 55,000 retracted or corrected over the same period — a ratio of about seven to one, on the sleuths’ own figures. Retractions passed 10,000 in a single year for the first time in 2023, and more than 8,000 of those came from one publisher’s cleanup of one imprint. A correction machine whose record year is dominated by a single accounting event is not a machine anyone can forecast from.

Frontier The audit that would replace the argument is cheap by the standards of the thing it would measure. The ERROR project pays up to 1,000 Swiss francs, about US$1,100, for a deep forensic review of a single paper. Three thousand papers at that rate is roughly US$3.3 million — under a tenth of the US$35–40 million one publisher booked in lost revenue in a single year from one integrity failure, and a normal size for a single research grant. The money is not the constraint.

Speculative The gain goes to the people who have to decide whether to trust a body of literature they did not produce. Meta-analysts, clinical guideline committees, funders choosing fields, and anyone assembling a training corpus all currently substitute a guess for a rate. It does not go to the screening vendors, whose product is priced against a threat that is more valuable to them unbounded than bounded, and that conflict is stated here rather than left for the reader to find. The condition on all of it is access: a blinded audit needs full text and submission metadata for a random sample, held by publishers who have never granted it, which is what section 5 is about.

Established This brief is one piece of arithmetic: the rate at which fabricated papers enter the literature against the rate at which the correction machinery removes them. Write the stock of uncorrected fraudulent papers at any moment as the running total of everything that got in minus everything that has been taken out. Entry is set by the production capacity of paper mills and the permeability of peer review. Removal is set by a chain with four links — detection, allegation, adjudication, notice — and the chain moves at the speed of its slowest link. Every quantity in that sentence is either measured badly or not measured at all, and the brief is about which is which.

Established A paper mill is a commercial operation that sells manuscripts, authorship slots on manuscripts, or both. Anna Abalkina traced 1,009 authorship offers posted on one Russian brokerage site between 2019 and 2021 to at least 434 papers, valuing the slots at US$6.5 million and finding authors in at least 39 countries — one mill, documented from its own advertising.

Frontier The distinction that organises everything here is between a screen and a prevalence. A screen returns the share of papers that trip a detector. A prevalence is the share that are actually fabricated. The two are connected by the detector’s sensitivity and false-alarm rate, and those are known only against training sets made of already-retracted papers — that is, against the output of the correction machinery whose adequacy is the question. The circularity is not a technicality. It is the reason the field’s headline numbers disagree by more than an order of magnitude.

Established Four briefs on this site touch this subject and none owns the flow. Scientific Governance Models owns the machines of self-governance and records that integrity enforcement runs two to four orders of magnitude behind the volume it polices; this brief takes that as given and asks what the volume is. Science of Science owns the production function of discovery. Pathological Science owns honest self-deception, a different failure with a different natural history. Forecasting Replication owns whether a published claim holds up, which is orthogonal: a fabricated paper can be perfectly replicable in the sense that nobody will ever try. The throughput question is unowned, and it decides whether the machinery is coping.

Speculative The framing under test is that self-correction is a process with a capacity, not a property science has. If that is right, the question is whether removal per year exceeds entry per year, and the literature cannot answer it in either direction. Handwave The stronger claim — that the stock will pass some threshold at which whole subfields become unusable — is asserted frequently, has no measured threshold attached, and is a conjecture rather than a finding.

Established Retractions passed 10,000 in a single year for the first time in 2023. Richard Van Noorden reported the count in Nature on 12 December 2023 from Retraction Watch data: more than 10,000 retraction notices, a record, with the retraction rate exceeding 0.2% of published papers and having more than trebled over the preceding decade. National rates per 10,000 articles, excluding conference papers, ran Saudi Arabia about 30, Pakistan 28.1, China 24.9 and Russia 23.5. Roughly 50,000 retractions were on record in total at that point. Established More than 8,000 of the 2023 notices came from Hindawi, the Wiley imprint, after a cleanup of compromised guest-edited special issues.

Frontier The 2023 figure kept moving after it was reported, which is itself a fact about the instrument. Guillaume Cabanac gave 2023 as “nearly 14,000” in Nature in August 2024, against fewer than 1,000 a year before 2009; a November 2025 bibliometric analysis counted more than 14,000 notices for 2023, more than 9,000 further retractions in 2024 and more than 5,000 in 2025 through August. Established The same analysis puts publication growth between 2000 and 2021 at 6.25% compound annually against 22.09% for retracted papers, with retraction rates per 10,000 of 31.97 for electrical engineering and computer science, 13.85 for clinical and life sciences, 11.04 globally and 3.25 for physics.

Established Retractions attributed specifically to paper mills are a much smaller number than total retractions, and they are concentrated to a degree that makes averages misleading. A 2026 Scientometrics study tracking the Retraction Watch database through 2024 identified 10,409 retracted paper-mill articles in total. China accounted for 9,137 of them, 87.8%; the next largest contributors were Ethiopia with 293, India with 156 and Pakistan with 86. The 2023 spike accounted for 6,780 of the total. Open-access articles were 80.8% of the set, hybrid 18.8% and subscription 0.4%. Frontier That geographic concentration is a fact about detected and retracted papers, not about produced papers, and every detector in section 5 has been challenged on exactly this point.

Established The journal-level distribution is extremely skewed. A 2025 cross-sectional study in Research Integrity and Peer Review identified 142 journals that retracted 2,051 paper-mill papers between January 2020 and December 2022. Half of them retracted exactly one; 24.6% retracted ten or more; four exceeded a hundred. Medicine and health made up 61.2% of indexed categories. Median editorial handling time was 86 days in the single-retraction group against 116 days in the heaviest group, the opposite of what the “fast journals are the bad journals” story predicts.

Established Removal is slow, and its speed is set by cleanup campaigns rather than by ordinary editorial process. The same 2026 tracking study gives a mean of 709.4 days between publication and retraction across the paper-mill set, rising from under 100 days before 2019 to roughly 980 days in 2021–2022 and falling to 602.6 days in 2023 — a fall the authors attribute to the Hindawi episode, that is, to one publisher retracting a large cohort at once rather than to the machinery getting faster.

Established The Hindawi episode is the only large-scale removal operation whose cost has been made public. Wiley reported US$9 million in lost revenue from a three-month special-issue pause, then an US$18 million quarterly decline and an expected US$35–40 million for the fiscal year. Nineteen Hindawi journals were delisted from Web of Science in March 2023 and four were shut as heavily compromised. Established Hindawi’s own account of the process is that at least two people, trained by the publisher’s staff, evaluated each paper against six named indicators of manipulation, with the per-paper assessment work outsourced to vendors.

Frontier On the entry side there are four families of estimate and they do not agree. The first is the publishing industry’s own: the 2022 COPE and STM paper mills research report is the source of the widely repeated figure that about 2% of all journal submissions across all disciplines originate from paper mills. It is (publishers’ own figures) and it is a submission-side number, not a published-literature number.

Frontier The second family is red-flag screening, and its headline number moved by a factor of five between preprint and publication. Bernhard Sabel and colleagues built a tallying rule from bibliographic metadata — a private rather than institutional email address, an international co-author, a hospital affiliation. The May 2023 medRxiv preprint analysed 15,120 PubMed-listed publications and reported red-flag rates rising from 13.4% to 33.7% in neuroscience and 19.4% to 24% in medicine between 2010 and 2020, with an overall 2020 figure of 28.8%; against about 1.33 million biomedical publications a year that extrapolation runs to hundreds of thousands of articles. Validation against 400 known fakes and 400 presumed non-fakes gave 86% sensitivity at a 44% false-alarm rate for the two-indicator rule. Established The peer-reviewed version, published in Naunyn-Schmiedeberg’s Archives of Pharmacology in 2025, reports 14.9% red-flagged in 2020 and 16.3% in 2023 — and, applying Bayes’ rule to a 94% sensitivity, 11.5% false-alarm rate and the resulting 33.5% positive predictive value, an estimated true fake rate of 5.8%, or about 107,800 articles a year. The authors set that against 5,671 biomedicine retractions in 2023, about nineteen times fewer.

Frontier That revision is the most instructive single event in this literature and it cuts both ways. The same team, the same method family and overlapping data produced a 2020 red-flag rate of 28.8% in preprint and 14.9% in print, and a headline fake rate five times lower once the positive predictive value was carried through. The published number is the better one. But the preprint figure is the one that entered public circulation, is still quoted, and is the source of the “a fifth of the literature is fake” framing. Established A screening rate is not a prevalence, and the arithmetic that converts one into the other needs a positive predictive value that in this case was 33.5% — that is, two out of three red-flagged papers were not judged fake.

Frontier The third family is machine text classification, and it is the largest screen anyone has run. Baptiste Scancar, Jennifer Byrne, David Causeur and Adrian Barnett screened 2,647,471 cancer papers published in PubMed between 1999 and 2024 with a BERT classifier fine-tuned on 2,202 retracted paper-mill papers against 2,202 controls. It flagged 261,245 papers, 9.87% with a 95% confidence interval of 9.83–9.90, as textually similar to retracted paper-mill work; more than 170,000 of those were affiliated with Chinese institutions, 36% of Chinese cancer output in the set. Internal validation gave accuracy 0.91, sensitivity 0.87 and specificity 0.96, with 39 false positives among 3,375 control papers. Established The authors state the limit themselves: flagging is a statistical screen, not an attribution of misconduct, the model may miss newer templates, and the ground-truth labels are the output of the correction machinery. João Phillipe Cardenuto has raised the specific objection that such a model may be learning patterns of scientific English written by Chinese authors rather than features of fabrication.

Frontier The fourth family is self-report, and it gives much smaller numbers for a narrower thing. Daniele Fanelli’s 2009 meta-analysis of 18 surveys pooled 1.97% of scientists admitting to fabrication or falsification, 95% confidence interval 0.86–4.45, against 14.12% reporting having observed it in colleagues. A 2022 Dutch national survey of 6,813 researchers using randomised response returned 4.3% for fabrication and 4.2% for falsification, with 51.3% reporting at least one frequent questionable practice. Established These are rates per researcher, not per paper, and a single mill-supplied author can carry many papers, so they bound the problem from a different direction rather than contradicting the screening numbers.

Established Independent forensic screens of specific evidence types land between the self-report and screening families. Elisabeth Bik, Arturo Casadevall and Ferric Fang inspected 20,621 papers from 40 journals published 1995–2014 by eye and found 3.8% containing inappropriate image duplication, with a forty-fold spread between journals, from 0.3% in the Journal of Cell Biology to 12.4% in the International Journal of Oncology. Of the 782 problematic papers, 45.5% involved repositioning and 25.1% alteration; those two categories are hard to produce by accident.

Established Where a screen has been run against a checkable ground truth, the error rate is high and the correction response is small. Pranujan Pathmendra, Yasunori Park, Francisco Enguita and Jennifer Byrne manually verified 6,647 nucleotide sequence identities in 334 screened Molecular Cancer papers and found 251 sequences, 3.8%, wrongly identified, affecting 91 papers — 27% of those screened, and 38% of the 2020 cohort. In Oncogene, 47 of 1,165 sequences were wrong across 21 of 42 papers, 40%. Sequence identity is objectively checkable; there is no judgement call. Established The correction response to the Molecular Cancer set was two corrections and four retractions, the retractions for other reasons, against 91 affected papers that had collectively been cited 7,932 times.

Established Removal does not stop citation. Carlos Candal-Pedreira and colleagues examined 1,182 retracted paper-mill papers in the BMJ in 2022 and found 93.8% had received at least one citation, median 11 with an interquartile range of 5 to 22 and a maximum of 131, with higher median counts in lower-impact journals; they note explicitly that they did not separate pre- from post-retraction citations. Established A 2026 Scientometrics study of 531 retracted articles in business, social sciences and technology did separate them, and found mean post-retraction citations of 9.80 in business against 9.10 before, 5.39 in social sciences against 2.27 before, and 6.96 in technology against 3.63. For two of the three fields, retraction was followed by more citation than the paper had attracted while it stood. Journal prestige and institutional reputation raised post-retraction citation; severity of the misconduct did not affect it.

Established The aggregate persistence number is large. Cabanac reported in August 2024 that some 62,000 annulled articles had by then been cited more than 836,000 times, and that the Problematic Paper Screener had flagged more than 16,000 papers carrying five or more tortured phrases, of which only 18% had been retracted. Frontier A 2025 systematic review of 24 studies reports 44.8% of paper-mill articles still being mentioned in systematic reviews in second-quartile journals — the laundering pathway David Bimler described in Nature: paper-mill products matter even when nobody reads them, because they are aggregated into reviews and enter the mainstream literature.

Established The correction machinery’s own throughput, where it has been counted, is very low. Mark Bolland and colleagues report in Nature that COPE member journals had published editorial notices for only about 6% — 11 of 192 — of publications known to contain uncorrected errors in one documented case set, and that of 153 reports submitted to the journal Anaesthesia, 44% were found to contain false data. Frontier The same authors record that nearly 80% of COPE’s 33 trustees and council members are currently or were recently employed in the publishing industry, which is relevant to how the 2% submission figure above should be read.

Established The people who do most of the detection are not employed to do it. Retraction Watch launched a paid sleuth-in-residence post in January 2025 at US$65,000 a year, funded by a single private donor and carrying defamation insurance — a detail that says as much about the work as the salary does. The ERROR project, described by Malte Elson in Nature in May 2024, pays reviewers up to 1,000 Swiss francs per paper with error bonuses to 2,500; it had sent almost 60 invitations, secured 13 agreements and completed one review at the time of writing. Established That is the entire funded, systematic, consent-based error-detection capacity the field had produced by mid-2024.

Established The census itself is private and incomplete. John Ioannidis and colleagues, working from 55,237 Retraction Watch entries as of 15 August 2024, could link 39,468 to Scopus records — about 71% — and found that 3.3% of 217,097 career-long top-cited scientists had at least one retraction, rising to 4.8–5.5% in clinical medicine and to 10.5% in complementary and alternative medicine. Frontier That nearly a third of the retraction record could not be matched to the main citation index is a measurement problem in its own right, and it is the reason every number here carries a source rather than an average.

Frontier The central open question is the prevalence of paper-mill output in a defined body of literature, and it is open because the instrument that would settle it has never been built. Everything in the record is either a count of what was caught, a screen whose positive predictive value is estimated against previously caught papers, or a survey of researchers about themselves. None of these is a prevalence, and combining them does not produce one.

Frontier Whether entry now exceeds removal is genuinely undetermined and the available numbers point both ways. Retraction volume grew at 22.09% compound annually between 2000 and 2021 against 6.25% for publication volume, which is the machinery catching up. But the 2023 record was dominated by one imprint’s cleanup, the 2024 count fell to more than 9,000, and the Sabel published estimate of about 107,800 biomedical fakes a year sits nineteen times above biomedical retractions for the same year. Speculative A reading in which detection has become efficient and the stock is now shrinking is consistent with the growth rates and inconsistent with the prevalence estimates; a reading in which the stock is growing is the reverse. The data do not adjudicate.

Frontier Nobody has measured the false-negative rate of any detection signal against real fabricated papers that were never caught. By construction this cannot be done without a ground truth independent of the correction machinery, which is the same missing object. Every reported sensitivity in this brief — 0.86 and 0.90 for the Sabel rules, 0.87 for the cancer classifier, the F-scores near 0.91 for image provenance analysis — is sensitivity against known positives, and known positives are papers that were caught.

Frontier Whether the geographic concentration in the retraction record is real or an artefact of detection is unresolved and consequential. China accounts for 87.8% of retracted paper-mill articles and 36% of flagged cancer papers, yet national retraction rates per 10,000 put Saudi Arabia at about 30 and Pakistan at 28.1 above China at 24.9. Frontier The competing explanation — that text classifiers trained on caught papers learn the English of non-native writers rather than the signature of fabrication — has been raised by a researcher who builds such detectors and has never been tested by holding language constant.

Speculative Whether large language models have changed the entry rate or only the texture of the output is open. The claim in circulation, reported in Nature in 2025, is that at least 10% of PubMed abstracts published in 2024 were written with a language model — a statement about writing assistance, not about fabrication. Frontier It bears on detection: tortured phrases are an artefact of paraphrasing tools that generative models have made obsolete, so a signal that found more than 16,000 papers should decay, and nobody has published a measurement of that decay.

Frontier The relationship between post-retraction citation and downstream contamination is measured at the citation level and not at the conclusion level. It is established that retracted papers keep being cited, and in two of three fields studied are cited more after retraction than before. Speculative What is not established is how often a fabricated result changes the conclusion of a meta-analysis, a clinical guideline or a review. The 44.8% figure for paper-mill articles appearing in second-quartile systematic reviews is an exposure number, not an effect number, and the effect number is the one that matters.

Speculative Whether correction capacity can scale without abandoning per-paper adjudication is the governance question underneath the arithmetic. The Hindawi operation resolved on the order of 8,000 papers with two trained evaluators each against a checklist, and was done by a publisher retiring an imprint it had bought. Handwave Cohort-level annulment — retracting a journal-year or a special issue wholesale on statistical grounds — is discussed in the field and has no published due-process framework, no false-positive budget and no test case outside an owner acting on its own property.

Established The binding technical constraint is not detection, it is access to the material detection needs. Every large screen in the record runs on what is public: titles, abstracts, metadata, and sometimes figures. The signals that discriminate best — duplicated referee reports, submission-system fingerprints, editorial timing, email and IP provenance — live inside publisher submission systems and have never been made available to an independent party for a random sample.

Established Tortured phrases measure paraphrasing, not fabrication. The signal, identified by Guillaume Cabanac, Cyril Labbé and Alexander Magazinov in 2021, is the appearance of synonym-substituted technical terms — “counterfeit consciousness” for artificial intelligence, “colossal information” for big data. What it detects is that text passed through a synonym tool, which is neither necessary nor sufficient for fabrication. Established Its practical limit is in the numbers: more than 16,000 papers flagged at five or more phrases, 18% retracted.

Established Image duplication measures figure reuse, and automating it has never been validated against an unknown-prevalence sample. Bik’s 3.8% came from human inspection of 20,621 papers. João Phillipe Cardenuto, Daniel Moreira and Anderson Rocha published a provenance-analysis method in PLOS ONE in 2024 that groups duplicated and manipulated regions across documents, reporting F-scores of 0.91 to 0.92 at document level on a 4,756-document set built from 121 known stock-photo mill articles plus distractors. Frontier That is performance on a set whose positives were chosen; the operating characteristic on a random field-year sample is unpublished.

Established Submission fingerprints measure duplication in the review process, and they are the publishers’ strongest signal because only publishers can see them. Adam Day showed in Scientometrics in 2022 that identical referee comments from nominally different reviewers are rare enough that duplication is itself informative, because mills create fake referee accounts and reuse reports; he is explicit that the method flags candidates for manual investigation rather than diagnosing misconduct. Frontier Day is chief executive of a paper-mill detection company, stated here because it is (vendor) work on a signal whose value cannot be independently checked without the data his customers hold. María Ángeles Oviedo-García documented the adjacent pattern in 263 review reports: boilerplate affirmations recycled across unrelated papers, combined with coercive citation.

Frontier Authorship-network structure measures collaboration shape and carries the field’s cleanest published false-positive comparison. Simon Porter and Leslie McIntosh identified 2,056 researchers between 2020 and 2022 whose egocentric networks had unusually low clustering coefficients alongside high output. Ten percent of tortured-phrase papers contained one of these authors directly and 37% contained a collaborator of one, against 1.2% with a standard deviation of 0.15 in random control samples. Frontier Their publisher-exposure figures — Hindawi about 4% in 2022, MDPI about 3%, most others 1–2% — derive from a commercial bibliometric database maintained by the authors’ employer, which is (vendor) provenance on the largest single set of numbers in this part of the field.

Frontier Citation cartels are the least measured of the four signals. The structure — groups of authors citing each other at rates far above network expectation — was formalised by Iztok Fister Jr., Iztok Fister and Matjaž Perc in Frontiers in Physics in 2016 as a graph-query problem, using a threshold of ten mutual citations in worked examples. Frontier That paper is a method proposal demonstrated on artificial networks; it reports no count of cartels found in real data. The 2025 bibliometric analysis attributes about 1,800 retractions between 2023 and 2025 to citation manipulation, which is a count of adjudicated cases, not a prevalence.

Established Predicting retraction from public metadata alone performs barely above chance. Aaron Fletcher and Mark Stevenson matched 4,514 retracted articles to 4,514 non-retracted controls from 2000–2020 and tested traditional classifiers against large language models; the best overall accuracy was 0.682, with publication year the single most important feature. Frontier A model whose strongest predictor of retraction is when the paper was published is describing the history of the correction machinery, not the properties of fraudulent papers, and that is the clearest available statement of how far metadata-only detection is from the job.

Established Adjudication, not detection, is the narrow link. Detection now operates at millions of papers: 2.6 million cancer papers screened by one classifier, about 60,000 submissions a month through the STM Integrity Hub duplicate checker, which flags around 1.5% as duplicates. Adjudication operates at the scale of two trained humans per paper. Established The Hindawi cleanup is the existence proof that the second can be scaled to thousands, and the US$35–40 million revenue figure is the only public price tag attached to doing so.

Established The first dependency is a sampling frame: a defined field-year from which a random sample can actually be drawn. Even the caught population is incompletely indexed — the retraction record matched Scopus for 39,468 of 55,237 entries. An audit needs a population definition that is closed, enumerable and independent of any detector.

Established The second is forensic review capacity priced and staffed. The two available price points are the ERROR project’s up-to-1,100 US dollars per paper and Hindawi’s outsourced two-evaluators-per-paper checklist, whose unit cost has never been published. Frontier Nobody has published a learning curve for forensic review, which is the number that decides whether an audit of several thousand papers is a three-month exercise or a three-year one.

Established The third is verifiable evidence types inside the sampled papers. The Byrne group’s nucleotide-sequence work is the model: an objectively checkable claim inside the paper, verifiable against public databases without contacting the authors, with a measured error rate of 3.8% of sequences and 27% of screened papers. Frontier Fields without such a claim type — most of the social sciences, much of engineering — would need a different forensic protocol, and no published protocol covers them.

Frontier The fourth is a calibration set of confirmed negatives, and it may not be obtainable. Sabel’s validation used 400 “presumed non-fake” papers and the authors say plainly that they cannot be certain those are not also fake; every sensitivity figure in this brief inherits that uncertainty. Speculative A cohort produced under observed, audited conditions and then tested blind is the only clean way to construct confirmed negatives, and no such cohort exists.

Established The fifth is data-sharing consent from publishers. Submission-system signals are the ones that discriminate, and the 2% submission estimate, the 1.5% duplicate rate and the referee-report duplication work all rest on data held privately by the organisations being measured. Frontier Publishers have pooled some of this through the STM Integrity Hub for their own screening; they have not released a random sample for independent audit, and no funder has made that release a condition of anything.

Established The sixth is a notice standard so that removal propagates. Retraction status has to travel to indexes, reference managers and downstream corpora, and the persistence numbers are evidence that it does not. Frontier That is an infrastructure dependency rather than a research one, closer to the metadata work in Metrology Infrastructure than to detection.

Frontier The decisive experiment is a blinded prevalence audit: a random sample of several thousand papers from one defined field and year, run through every published detection signal and then through independent human forensic review, producing a prevalence estimate with stated bounds. It is decisive because every disagreement in this brief reduces to the absence of that one number, and because a single well-run audit would convert four incomparable instruments into four calibrated ones.

Established The design is specified by what already exists. The sampling frame is a field-year enumerable from a public index — the cancer literature screen used 2.6 million PubMed records from 1999 to 2024 and demonstrates that enumeration is possible. The sample size is set by the width of the interval wanted: at a prevalence near 5%, a sample of 3,000 gives a 95% interval of roughly plus or minus 0.8 percentage points, which is narrow enough to separate the 2% industry figure from the 5.8% Bayes-corrected estimate and both from a 10% screen rate.

Established The blinding is the part that does not exist anywhere in the current record. Reviewers must not know whether a paper was flagged, must not know the journal, and where possible must not know the author country, because the country-versus-language confound is the most contested inference in the field. Frontier Redacting country while preserving enough context for forensic judgement is unsolved and would need piloting.

Established The forensic protocol has published components. Nucleotide-sequence verification against public databases, as applied to 6,647 sequences across 334 papers; image duplication inspection at the three-category standard used on 20,621 papers; reagent and cell-line identity checks; and internal numerical consistency. Frontier What is missing is an adjudication rubric with published inter-rater agreement: Hindawi used six indicators and two trained evaluators per paper and published neither the agreement statistics nor the disagreement rate.

Frontier The audit should be a registered report with the analysis plan fixed before any paper is opened. A headline that moved from 28.8% to 14.9% red-flagged and then to 5.8% estimated prevalence between preprint and print is exactly the pattern pre-registration exists to prevent, and it happened in the field’s most-cited estimate.

Established A second experiment is cheaper and is the one a publisher could run tomorrow. Take a single journal-year, apply every signal, refer every flagged paper and a matched sample of unflagged papers to the same blinded forensic panel, and publish the two-by-two table. That table — flagged and confirmed, flagged and cleared, unflagged and confirmed, unflagged and cleared — is the positive predictive value and the false-negative rate that nothing in the published record currently supplies. Frontier The STM Integrity Hub screens about 60,000 papers a month and could generate the flagged arm of that table as a by-product.

Frontier A third experiment settles the language confound and needs no new data. Hold author country constant and vary language quality, or hold language quality constant and vary country, in the training and test sets of an existing text classifier, and report the change in flag rate. Established The materials exist: 2,202 retracted paper-mill papers and 2,202 controls were used to train the cancer classifier, and the objection that it may be learning Chinese-authored scientific English rather than fabrication was raised publicly at the time of publication.

Established A natural experiment is already running and its result will arrive without funding. If the tortured-phrase signal decays while red-flag and network signals hold, the mills changed tooling; if all decay together, detection is losing. Frontier The Problematic Paper Screener has published flag counts continuously since 2021, so the time series exists and nobody has published the decay analysis.

Speculative The experiment that would settle downstream harm is larger and nobody has proposed a tractable version. It would take a set of systematic reviews and clinical guidelines, remove every flagged input, recompute the pooled estimates and report how often the conclusion changes. Handwave Doing this at the scale of a whole evidence base assumes an accurate flag on every input, which is what the prevalence audit is for; run first, it would measure the detector rather than the contamination.

Established The audit is an operations problem before it is a research problem. Three thousand papers through a blinded panel at two independent reviewers each is 6,000 reviews plus adjudication of disagreements. At the ERROR project’s rate of up to about US$1,100 per deep review, reviewer cost alone runs from US$3 million to US$6.6 million before coordination, redaction and statistical support.

Established Redaction is the hardest piece of engineering in the design. Stripping journal, author, affiliation and country from a full-text PDF while preserving methods, figures and sequences is not a solved pipeline; funding acknowledgements, reagent suppliers and language register all leak the attributes the blinding is meant to remove. Frontier No published audit has attempted it, so the residual leakage rate would have to be measured in the pilot.

Established The detection stack itself is buildable from published components, and the missing piece is not a component. Text classification at the 2.6-million-paper scale, image provenance analysis near F-score 0.91, authorship-network anomaly detection and tortured-phrase screening have all been published. Frontier What has never been published is a fused score with a calibrated operating point, because calibration requires the prevalence the audit would supply.

Speculative Cohort-level annulment is the only proposed engineering answer to the adjudication bottleneck and it has no safety case. The Hindawi operation is the closest thing to a demonstration: thousands of papers resolved against a six-indicator checklist with two evaluators each, executed by the owner of the journals. Handwave Generalising that to an independent body annulling a journal-year it does not own requires a due-process standard, a false-positive budget and an appeal route, none of which has been drafted. The governance machinery that would have to hold such a power is the subject of Scientific Governance Models, which finds that the interventions with demonstrated effects are checkpoints applying a rule to every item — and cohort annulment is exactly that, applied at the wrong end of the pipeline.

Frontier Automating the forensic panel is the obvious temptation and would destroy the experiment. The audit’s value is that its ground truth is independent of the detectors being calibrated; a language model as reviewer reintroduces the circularity in a less visible form. Speculative A defensible hybrid is machine triage of objectively checkable items with human judgement reserved for the fabrication call, and even that needs its agreement with unaided human review reported.

Frontier Generative language models sit on both sides of this ledger and the net sign is unknown. On the production side they remove the last cheap signal: text that reads naturally does not carry tortured phrases. On the detection side they make large-scale semantic screening affordable. Established The measured facts available are that a BERT-class model screened 2.6 million papers at accuracy 0.91 on its own validation set, and that metadata-only retraction prediction with language models topped out at 0.682 accuracy.

Established Image forensics is the most mature adjacent technology and the one with the longest measured record. Human inspection established 3.8% duplication in 20,621 papers in 2016; automated provenance analysis reached F-scores of 0.91 to 0.92 at document level in 2024. Frontier Commercial image-screening tools are now standard in publisher workflows and their operating characteristics on random samples are (vendor) figures, not published ones.

Established Persistent identifiers are the infrastructure that makes both the fraud and the audit possible. The authorship-network method depends on resolving the same person across papers, and the Retraction Watch to Scopus linkage reached only 39,468 of 55,237 records. Frontier An identifier system that can be bought is a signal that can be spoofed.

Established Post-publication peer review is the detection layer that actually works and has no institutional standing. PubPeer carried comments on 191,463 articles by August 2024. Frontier Its throughput is bounded by unpaid volunteer time and by legal exposure — the paid sleuth post comes with defamation insurance because that is a real cost of the work.

Frontier Bug-bounty economics is the nearest working analogue and the transfer is unproven. Malte Elson’s comparison in Nature is explicit: Google paid out US$10 million for vulnerabilities in 2023, against an error-detection field with essentially no paid capacity. Established ERROR’s throughput at launch — almost 60 invitations, 13 agreements, one completed review — says consent is the rate-limiting step rather than money.

Frontier Automated research pipelines change the question rather than answering it. Artificial Scientists owns the verification problem for machine-generated results, and Cloud Laboratories owns remotely executed experiments whose logs exist by construction. Speculative An execution record that cannot be fabricated after the fact would move fabrication detection from textual forensics to log audit, which is a different and probably easier problem — but only for work done that way, which is a negligible share of the literature this brief is about.

Established No institution currently has both the mandate and the data to run a prevalence audit. Publishers hold the submission data and are the measured party. National integrity offices have statutory authority over funded researchers and no access to the literature as a whole. Indexers hold the corpus and no investigative mandate. Sleuths have the skills and neither authority nor funding. Frontier The audit is therefore an institutional design problem that happens to have a measurement at the end of it.

Established The industry’s own bodies are staffed by the industry, and this is documented rather than alleged. Nearly 80% of COPE’s 33 trustees and council members are currently or recently employed in publishing, on Bolland and colleagues’ count in Nature; COPE has more than 14,000 member journals, and in one documented set its member journals issued editorial notices for 11 of 192 publications known to contain uncorrected errors. Frontier Whichever way that 6% is explained, the 2% submission-prevalence figure that circulates as the industry baseline comes from the same institutional setting and should be read as (publishers’ own figures).

Established The correction record is a private asset. The continuous census of retractions is the Retraction Watch database, distributed via a Crossref-hosted repository that Retraction Watch’s own site points readers to, rather than a public register maintained by any government or indexer. Frontier Open distribution is a real improvement and does not change the point: the authoritative count of the literature’s failures is compiled by a small non-profit newsroom whose senior detection post is funded by one private donation of US$65,000 a year.

Established Coordination bodies exist and their output is process rather than measurement. A cross-sector effort bringing together funders, publishers and research organisations to tackle paper mills was launched in January 2024 and reported in Nature; the STM Integrity Hub pools screening across publishers and runs a duplicate-submission check over about 60,000 papers a month. Frontier Neither has published a prevalence estimate, and pooled screening by the measured parties is not an audit of them.

Frontier The obvious obstacle is that the result embarrasses everyone. A credible prevalence number either vindicates the industry’s 2% or supports something closer to the screening rates; the first devalues a growing integrity-services market and the second indicts the journals that published the papers. Speculative That both outcomes are costly to a party whose cooperation is required is the most plausible explanation on offer for why an audit costing a few million dollars has not been run, and it is a conjecture about motives rather than a documented refusal.

Established Enforcement capacity is covered elsewhere on this map and constrains everything here. Scientific Governance Models records a statutory integrity office publishing single-digit annual findings against a year of more than 10,000 retractions. Frontier This brief adds one observation: the removal that did happen at scale was executed by a commercial publisher acting on its own property for commercial reasons, which is not a governance mechanism that generalises.

Established The dominant ethical hazard is a false positive attached to a named person. Every published detector reports a substantial false-alarm rate: 44% for Sabel’s two-indicator rule, 11.5% for the tallying rule with a positive predictive value of 33.5%, and 39 false positives among 3,375 control papers for the cancer classifier. The BMJ authors state the principle directly — flagging is a statistical screen, not an attribution of misconduct — and it is the sentence that most often goes missing when these numbers are reported.

Frontier The country confound turns a measurement error into an accusation against a nationality. China accounts for 87.8% of retracted paper-mill articles and 36% of flagged cancer papers; the objection that text classifiers may be learning the English of Chinese-affiliated authors rather than the signature of fabrication was raised by a detection researcher and has not been tested. Established The experiment that would test it is cheap and is specified in section 7; that it has not been run while the flag rates are quoted publicly is a choice with consequences for individuals.

Established The people doing the detection carry personal legal risk. That an employer must insure a sleuth against being sued for identifying fabricated data describes the incentive structure precisely: the cost of accusing falls on the accuser and the cost of publishing falls on nobody.

Frontier Patients are the population most exposed and the exposure is measured only at the citation level. Medicine and health made up 61.2% of categories among journals retracting paper-mill work, and 44.8% of paper-mill articles were still being cited in second-quartile systematic reviews. Speculative How many clinical recommendations rest on fabricated inputs is unknown, and section 7 says why nobody has proposed a tractable way to find out.

Frontier The incentive to buy authorship is created by the institutions that evaluate researchers, and the market price is documented. One Russian brokerage’s slots across 2019 to 2021 were valued at US$6.5 million across 1,009 offers, with authors in at least 39 countries. Established Hospital affiliation is one of the three red-flag indicators in the Sabel rule, which is a statement about clinician promotion criteria as much as about fraud. Scientific Funding Models owns the evaluation-design question; this brief notes only that demand is institutional and supply responds to it.

Speculative A published prevalence number would itself be a political object. A credible figure of 5% for a national literature is a diplomatic event; a figure of 20% is a different one. Handwave Arguments that the audit should therefore not be run, or should be run without publishing country breakdowns, are made informally in this field and have no worked-out version; the position taken here is that an unmeasured rate is already being used politically, with worse evidence.

Established The literature is a shared capital stock and this brief is about its depreciation rate. At least 400,000 papers from 2000 to 2022 carry paper-mill hallmarks on the sleuths’ count, against 55,000 retracted or corrected. Whatever the true prevalence, the removal side is small relative to the flagged side, and the gap has been accumulating for two decades.

Frontier The mechanism of harm is laundering rather than readership. David Bimler’s formulation in Nature is that paper-mill products are a problem even if nobody reads them, because they are aggregated into review articles and laundered into the mainstream literature. The measured support is the 44.8% figure for paper-mill articles in second-quartile systematic reviews and the post-retraction citation means that exceed pre-retraction means in two of three studied fields.

Speculative Machine training corpora are the newest transmission path and the least examined. A fabricated result that enters training data is not cited, is not retractable and leaves no trace in the citation graph. Handwave That this is a qualitatively new contamination channel is plausible and unmeasured: no published study estimates the share of any major training corpus that is retracted or flagged work.

Frontier The concentration of the problem is the most hopeful fact available and it is not established. Half of the 142 journals that retracted paper-mill papers retracted exactly one, while four exceeded a hundred; Bik found a forty-fold spread between journals; 87.8% of retracted paper-mill articles trace to one country. Speculative If production is genuinely concentrated in a small number of mills and venues, the correction problem is tractable at a few thousand papers a year rather than a few hundred thousand. If the concentration is an artefact of where detectors have been pointed, it is not.

Frontier The reputational cost is already being paid by fields rather than by papers. Micro and long non-coding RNA carried 3,123 retractions in the 2025 topic analysis and human-computer interaction 2,137. Speculative Whether readers discount work from high-retraction fields, and by how much, has not been measured.

Established Self-correction has been demonstrated at scale exactly once, and commercially. The 2023 record year was one publisher removing about 8,000 papers from an imprint it had bought, at a cost it put at US$35–40 million in lost revenue. Frontier That is evidence the literature can be cleaned and it is evidence the cleaning happens when a commercial owner decides to absorb the cost. No comparable operation has been executed by a scholarly society, a funder or a state.

These horizons track the measurement of the throughput problem, not the volume of fraud itself — the second cannot be forecast until the first exists.

• 10 yr: Frontier Either a blinded prevalence audit of one field-year is published and the 2% to 29% spread collapses to an interval, or it is not and the field continues to argue from incomparable screens. The cheap version — a single journal-year with a published two-by-two table of flagged-and-confirmed against unflagged-and-confirmed — is within any large publisher’s existing capability and requires no new science. The tortured-phrase signal decays as generative models displace synonym tools, and whether the other signals decay with it is settled by the time series that already exists.
• 25 yr: Speculative Either retraction and correction status propagates reliably into indexes, reference managers and training corpora — the plumbing problem described in Metrology Infrastructure — and the 836,000 citations to annulled work becomes a historical figure, or post-retraction citation remains the normal case and removal continues to mean removal from one database. Cohort-level annulment either acquires a due-process standard or remains a thing only an owner can do to its own property.
• 50 yr: Speculative The prevalence question is either a routine audited statistic reported per field and year, like a national accounts revision, or it has been abandoned in favour of provenance: work that carries an execution record from instruments and automated laboratories, as described in Cloud Laboratories, against work that does not. The second path does not clean the existing corpus; it partitions the future one.
• 100 / 250+ yr: Handwave Any statement at this range is about whether a text-based, venue-mediated scientific literature still exists as the unit of record. If it does not, the throughput problem is replaced by whatever verification attaches to the successor artefact, and nothing in the current evidence base constrains what that is. This is placed here as a handwave rather than omitted because the alternative is to imply the current arrangement is permanent.

• Depends on the governance and enforcement record owned by Scientific Governance Models, which supplies the enforcement-capacity numbers this brief treats as given, and the bibliometric measurement apparatus discussed in Science of Science and Research Front Detection, whose validation traps — detectors tuned backwards on known positives — are the same traps that appear here. No result on this map blocks the prevalence audit: it is an operations and access problem, not a scientific one.
• Requires (not on this map) a blinded prevalence audit of a defined field-year, because every quantity in this brief is a screen or a count and neither is a rate; full text and submission metadata released for a random sample by the publishers who hold them, since the discriminating signals are the ones only they can see; a public register of retraction and correction notices, because the authoritative census is currently a non-profit newsroom’s database and a third of it does not match the main citation index; forensic review capacity at thousands of papers a year, against a demonstrated funded capacity of one completed independent review at the ERROR project’s launch; and a buyer willing to pay for a prevalence number with bounds rather than for a screening subscription priced against an unbounded threat.
• Enables calibration of every existing detector to a known operating point; honest contamination estimates for systematic reviews, clinical guidelines and training corpora; a defensible basis for cohort-level annulment or for refusing it; and evaluation policy that can be argued from a rate rather than from an anecdote.
• Adjacent to Pathological Science for the honest-error failure mode this brief deliberately excludes, Forecasting Replication for whether a published claim survives testing, Information Integrity for the same production-to-belief chain outside science, Artificial Scientists for verification of machine-produced results, and Scientific Funding Models for the evaluation incentives that create demand for bought authorship.

Established “Retractions passed 10,000 in 2023, so fraud is now running at about 10,000 papers a year.” The retraction count is a measure of the correction machinery’s output, not of the entry rate. More than 8,000 of the 2023 notices came from one publisher clearing one imprint; 2024 fell back to more than 9,000. A count dominated by a single accounting decision cannot be read as an incidence.

Frontier “About a fifth of the biomedical literature is fake.” This traces to the Sabel group’s 2023 preprint, which reported 28.8% red-flagged in a 2020 sample. The peer-reviewed version reports 14.9% red-flagged for the same year and, carrying through a positive predictive value of 33.5%, an estimated true fake rate of 5.8% and about 107,800 articles a year. Established The number that entered circulation was the screening rate from the preprint; the number the authors now defend is roughly five times lower.

Established “Machine screening found 10% of cancer papers to be fake.” It found 9.87% textually similar to retracted paper-mill papers, with the authors stating that flagging is a statistical screen and not an attribution of misconduct, that the model may be detecting one template family, and that the ground-truth labels are themselves the output of the correction machinery. Frontier A published objection holds that such a model may be learning Chinese-authored scientific English rather than fabrication, and the experiment that would separate the two has not been run.

Established “Peer review catches fabrication.” The signal that best distinguishes mill-produced submissions is duplication in the referee reports themselves, because mills supply the reviewers too; more than 6,400 retractions between 2023 and 2025 were attributed to fraudulent peer review.

Established “Retraction removes a paper from the literature.” It does not remove it from the citation graph. Some 62,000 annulled articles had been cited more than 836,000 times by August 2024; 93.8% of 1,182 retracted paper-mill papers had received at least one citation; and in a 2026 study of three disciplines, mean post-retraction citations exceeded pre-retraction citations in social sciences and technology. Frontier Severity of the misconduct had no significant effect on post-retraction citation in that study, while journal prestige raised it.

Frontier “Open access caused this.” Open-access articles are 80.8% of retracted paper-mill papers, which looks decisive until the denominator is considered, and the damage was concentrated in guest-edited special issues rather than in open access as such. Established Open access was 72.6% of journals with a single paper-mill retraction against 19.2% of journals with ten or more, which runs against the simple story.

Established “The sleuths are cranks with an axe to grind.” The detection record is checkable: 3.8% image duplication across 20,621 papers with a published category scheme; 251 of 6,647 nucleotide sequences wrongly identified in papers whose sequences anyone can look up; 1,009 authorship advertisements traced to at least 434 papers. Frontier The relevant criticism is the opposite one — that detection is concentrated where sleuths have chosen to look, which biases the map of where fraud is.

Speculative “Artificial intelligence will solve detection.” The best published metadata-only retraction prediction reaches 0.682 accuracy, with publication year as the most important feature. Text classification at scale reaches 0.91 accuracy against its own training distribution and inherits that distribution’s biases. Handwave The assertion that a sufficiently large model will identify fabrication without a calibration set is the step where this argument works by assertion: without a known prevalence there is nothing to calibrate against, and the model’s confidence is not evidence.

Established “Publishers have strong incentives to clean up.” The one large cleanup on record cost its publisher US$35–40 million across a fiscal year, plus 19 journals delisted and four closed. Frontier That is the measured price of correcting the record at scale, and it was paid by a firm that had bought the imprint and chose to retire the brand.

Handwave “Science is self-correcting, so this sorts itself out.” Self-correction is a process with a throughput, and the throughput numbers in this brief are: a mean of 709.4 days from publication to retraction in the paper-mill set; editorial notices for 11 of 192 known-flawed publications in one documented case set; two corrections and four retractions against 91 papers with verifiably wrong sequences; 18% retracted among more than 16,000 papers carrying five or more tortured phrases; one completed independent review at the launch of the only funded error-detection programme. Established None of those is a claim that correction never happens. Each is a rate, and the argument that the rate is adequate has never been made with numbers.