A rising share of retractions traced to organized fraud rather than individual error share two tell-tale signatures: manuscripts produced by paper mills, and prose containing tortured phrases — synonym-mangled versions of standard terminology that betray machine-paraphrased or disguised-plagiarized text. Neither signal proves misconduct on its own, but together they are now central to how publishers and institutions triage suspect submissions before they ever reach peer review.
This guide covers what each signal actually is, how they’re detected in practice, why they often — but not always — travel together, and what the publishing industry’s two main coordinated responses (the STM Integrity Hub and COPE’s 2025 retraction guidelines) actually do.
What a paper mill is
CASRAI’s Dictionary defines a paper mill as a commercial operation that produces fabricated or low-effort manuscripts and sells authorship slots, or that brokers acceptance of such manuscripts into journals. The defining feature isn’t poor writing quality — it’s the commercial, for-profit production of research outputs designed to look like real science without an underlying real study. Common patterns documented by publishers and integrity researchers include:
- Templated manuscripts. Batches of papers that share an identical structural skeleton — the same figure layout, the same statistical tables, the same discussion-section phrasing — with only the gene name, disease, chemical compound, or dataset swapped out.
- Authorship for sale. Slots on an already-accepted or in-preparation manuscript offered to buyers who had no role in the underlying work, sometimes advertised openly on social media or via brokers.
- Reused or recycled images. Figures — particularly Western blots and flow cytometry plots in biomedical literature — reappearing, sometimes digitally altered, across manuscripts submitted to different journals under different author names.
- Compromised peer review. Fabricated reviewer identities or reviewer-suggestion systems gamed so a paper mill effectively reviews its own submission.
COPE and STM’s joint 2022 research report (conducted with Maverick Publishing Services, based on real publisher-submitted data) was the first industry-wide attempt to size the problem using shared data rather than any single publisher’s anecdotal experience, and it concluded the scale required coordinated, cross-publisher action rather than each journal fighting it alone — the direct origin of the United2Act initiative covered below.
Not every low-quality or templated-looking manuscript is paper mill output. Legitimate contract research organizations, professional medical-writing services operating under recognized authorship guidelines (e.g. GPP 2022 for industry-sponsored publications), and English-language editing services that improve a manuscript without altering its scientific content are not paper mills — the distinguishing question is whether the underlying research and the claimed authorship contributions are real.
Tortured phrases as a detection signal
A tortured phrase is an unusual paraphrase of established technical terminology, produced by automated synonym substitution, that a domain expert immediately recognizes as a corrupted version of a standard term. The concept and term were introduced by Guillaume Cabanac (Université de Toulouse), Cyril Labbé (Université Grenoble Alpes), and Alexander Magazinov in their 2021 paper “Tortured phrases: A dubious writing style emerging in science” (arXiv:2107.06751), which documented the pattern across established journals and was covered by Nature the same year.
The mechanism is straightforward: paraphrasing software — used either to evade plagiarism-detection tools by substituting synonyms for lifted text, or to disguise machine-translated or AI-generated text as original writing — has no domain knowledge, so it substitutes synonyms for fixed technical terms that should never be paraphrased. The result is text that reads as plausible English at a glance but is instantly recognizable as wrong to anyone who knows the field. Documented, widely-cited examples include:
- “counterfeit consciousness” for artificial intelligence
- “profound learning” for deep learning
- “colossal information” for big data
- “cruel temperature” for mean temperature
- “irregular esteem” for random value
Cabanac and colleagues built the Problematic Paper Screener, a public tool that queries a database of tens of millions of indexed articles against a list of thousands of known tortured phrases to flag papers for post-publication reassessment — the same underlying detection logic several commercial and publisher-run screening tools now license or replicate at submission time, before publication rather than after.
Two caveats matter for anyone using this as a signal rather than a definition. First, a tortured phrase is evidence of paraphrasing-tool use, not proof of fabricated results on its own — genuinely real research can, in principle, be reported using tortured phrasing if an author (or an editing vendor working for them) ran the text through the wrong kind of tool. Second, genuinely new terminology introduced in good faith by an emerging subfield, properly defined at first use and applied consistently thereafter, is not tortured phrasing — the tell is a term that reads as a corrupted synonym for something that already has an established name, not an unfamiliar but internally consistent new one.
Why the two signals travel together
Paper mills and tortured phrases are not the same thing, but they correlate for a structural reason: a mill producing large volumes of templated manuscripts across many buyers has a strong incentive to make each output look textually distinct enough to survive plagiarism-similarity checks like Turnitin or iThenticate, which compare submitted text against prior published text. Running a shared template through automated paraphrasing software is a cheap way to generate many surface-distinct versions of the same underlying boilerplate — which is exactly the kind of text tortured-phrase detectors are built to catch. That’s why STM Integrity Hub-connected tools and several individual publisher pipelines now run tortured-phrase screening as one signal alongside citation-pattern analysis, reused-image detection, and submission-metadata checks (shared IP addresses, unusual co-authorship networks, disposable-looking author email domains) rather than relying on any single indicator.
Neither signal is used in isolation to reject or retract a paper. COPE’s own case guidance and the flowchart-based decision process described in CASRAI’s COPE flowcharts guide treat a tortured-phrase hit or a templated-manuscript pattern as grounds to investigate, not as automatic proof of misconduct — editors are expected to look at the pattern in context (does the paper also show recycled figures, an implausible submission-to-acceptance timeline, or an author group with no verifiable institutional affiliation?) before acting.
Other red flags editors and integrity offices watch alongside these two
Paper mill and tortured-phrase detection is usually one part of a broader screening pattern. Related signals that frequently appear alongside them:
- Image manipulation — reused, duplicated, or digitally altered figures, especially Western blots and microscopy images.
- Citation cartels — coordinated groups of authors or journals that cite each other’s work disproportionately to inflate metrics.
- Fabrication and falsification of underlying data, of which paper mill output is a specific commercial variant.
- Submission from a predatory journal or via a compromised guest-editor special issue, a channel paper mills have specifically been documented exploiting because peer review is weaker or effectively absent.
- Sudden, statistically anomalous spikes in submissions from a single institution, geographic region, or research topic that don’t match that group’s known publication history.
What publishers and institutions are doing about it
The STM Integrity Hub
The STM Integrity Hub, operated by STM Solutions (the operational arm of the International Association of STM Publishers), is a shared, cloud-based screening platform that lets participating publishers check submitted manuscripts against pooled integrity signals before publication — an “early warning system” run collectively rather than by any one publisher alone, in a protected environment designed to respect data-privacy and competition-law constraints between competing publishers. It integrates a modular set of independent detection tools, including tortured-phrase screening, duplicate-submission and citation checks, reused-image detection (via tools such as Clear Skies’ Papermill Alarm), AI-generated-text detection, and cross-publisher submission-pattern analysis. As of a December 2025 report in Science Editor, roughly 40 publishers were using the Hub, screening over 125,000 papers a month and intercepting approximately 1,000 suspected paper mill submissions monthly across about 20 connected detection tools and seven editorial systems.
United2Act
United2Act is the multi-stakeholder initiative that grew directly out of the 2022 COPE/STM paper mill research: at a May 2023 summit, roughly 40 participants — publishers (including Elsevier, Springer Nature, Taylor & Francis, Wiley, and others), research institutions, funders, and infrastructure organizations (Crossref, ORCID, Clarivate, Digital Science) — agreed a consensus statement covering five areas of collaborative action: education and awareness, improving post-publication correction processes, researching paper mills, developing shared trust markers, and strengthening cross-sector communication. It is a coordination framework rather than a detection tool — its output is shared standards and practices that individual publishers and platforms (including the Integrity Hub) then implement.
COPE’s 2025 retraction guidelines
COPE’s Committee on Publication Ethics revised its retraction guidelines in September 2025 (Version 3, replacing the 2019 version), naming paper mill involvement, compromised peer review (fake reviewers, citation manipulation), unverifiable authorship, and undisclosed AI use as explicit retraction grounds for the first time — codifying paper mill activity as a named category rather than leaving it to fall under a more general “fraud” heading. The revised guidelines also recommend crediting third parties (post-publication reviewers, “sleuths,” or readers) who raised a valid concern in the retraction notice, with their permission, and require mass-retraction notices tied to coordinated fraud to state plainly that the retraction is part of a systematic pattern rather than an isolated case. See CASRAI’s guide to how a retraction actually happens for the full editorial process these grounds feed into, and the COPE flowcharts guide for how editors work through a suspected-misconduct case step by step.
What this means in practice for authors and research institutions
For legitimate authors and institutions, the practical takeaway isn’t defensive — none of these detection systems are designed to catch honestly reported research, including research that happens to use unfamiliar terminology or was run through a professional (non-paraphrasing) editing service. The institutional relevance is upstream of any individual manuscript:
- Research-integrity offices increasingly need to be able to explain paper mill and tortured-phrase findings to faculty and administrators when a retraction or expression-of-concern notice names an affiliated author, including cases where an author unknowingly bought into, or was added to, a paper mill product.
- Hiring, tenure, and grant-review committees relying on publication counts or citation metrics are exposed to the same fabricated-output problem — a paper mill manuscript that hasn’t yet been caught still counts toward an author’s record until it’s retracted.
- Institutions co-authoring or collaborating internationally should be aware that paper mill activity is not evenly distributed globally, and that a sudden, unexplained surge in a collaborator’s publication output is itself a signal worth a direct conversation, not an accusation.
Frequently asked questions
Is a tortured phrase proof that a paper is fraudulent?
No. It’s evidence that some portion of the text passed through automated paraphrasing software, which is a strong reason to investigate further — checking for other signals like recycled figures, an implausible submission history, or unverifiable authorship — but not, by itself, proof of fabricated data or a paper mill origin.
Can legitimate research accidentally trigger these detectors?
It’s possible but uncommon. A non-native-English-speaking author using an inappropriate machine-translation or paraphrasing tool (rather than a professional editing service) on an otherwise genuine manuscript could introduce tortured phrasing without any underlying misconduct. This is exactly why COPE guidance treats a detection hit as a trigger for editorial investigation, not an automatic rejection.
Who actually runs the STM Integrity Hub — is it a single company’s product?
It’s operated by STM Solutions on behalf of participating STM Association member publishers, and works by integrating multiple independent third-party and publisher-built detection tools (rather than being one proprietary detector) inside a shared, access-controlled environment.
Does using CRediT or transparent authorship statements prevent paper mill activity?
Not by itself, but it raises the cost of it: a paper mill selling authorship slots depends on contributor claims that can’t be checked. Requiring a specific, per-role CRediT statement gives editors a concrete, falsifiable claim to question if a listed author can’t describe their own contribution — one input among several, not a standalone defense.
What should an editor or institution do if they suspect a submission is paper mill output?
Follow the applicable COPE case guidance and flowchart rather than acting unilaterally — see CASRAI’s COPE flowcharts guide for the plagiarism/duplicate-submission/authorship decision trees, and the retraction process guide for what happens if the investigation confirms the concern after publication.
