Tortured Phrases in Medical Research: What the October 2026 Retraction Wave Reveals for Authors

When IOP Publishing processed a batch of post-publication integrity checks in October 2026, the results included phrases like "cerebral torment" in a paper that should have said "headache," and "armada the board" in one that meant "fleet management." The 80 or so papers retracted that week joined a running total that Retraction Watch, citing researcher Guillaume Cabanac, reported had passed 1,000 retractions from IOP for this specific problem. The cause in every case was the same: automated paraphrasing software that replaces technical terms with approximate synonyms, producing text that evades standard plagiarism detection but is recognizable as machine-generated to any reader who looks at it carefully.

For medical researchers, this might seem like someone else's problem. Physics and engineering conference proceedings are where most of these retractions have concentrated. That impression is wrong. The same paraphrasing tools, and the same underlying incentive to produce publishable text without original scholarship, have been documented in biomedical literature too. Publishers including MDPI and Hindawi have processed retractions for similar reasons. And the automated detection systems now deployed across major publishers do not limit their attention to non-medical fields.

What Tortured Phrases Actually Are

The term "tortured phrases" was coined by Guillaume Cabanac, Cyril Labbé, and Alexander Magazinov in a 2021 paper that described a systematic pattern in scientific publishing fraud. The phrases arise when someone runs a scientific text through an automated synonym-substitution tool, also called a paraphrasing spinner, in order to disguise either plagiarized text or machine-translated content as original writing.

The tools work by replacing words with synonyms from a general-purpose thesaurus. They have no knowledge of domain-specific technical terminology. "Artificial neural network" becomes "fake neural organization." "Breast cancer" becomes "bosom peril" or "bosom disease," depending on which synonym the spinner selects. "Deep learning" becomes "profound learning." "Random forest" becomes "irregular woodland." "Convolutional neural network" becomes "convolutional brain organization." The results read as strange to a human expert immediately, but they pass standard similarity-detection tools like iThenticate because the word strings no longer match the original source text.

The phenomenon is most dense in fields with high-volume conference proceedings, where editorial oversight has historically been lighter than in peer-reviewed journals, and where papers are sometimes submitted from regions where English is not the primary research language. But the tools themselves are freely available online, and their use in full journal article preparation, including medical research, has been documented repeatedly. Cabanac's Problematic Paper Screener, a public tool that searches approximately 130 million papers weekly, maintains a phrase list currently containing more than 7,900 confirmed pairs.

Not all of these phrases come exclusively from spinner tools. Some arise from non-native English speakers relying heavily on machine translation without reviewing the output carefully. A researcher who drafts in Mandarin, runs the text through an automated translator, and submits without expert editing may end up with phrasing that resembles tortured phrases even without any intent to deceive. This distinction matters when publishers investigate flagged papers, but it does not change the detection outcome, and it does not change the fact that the manuscript contains text that could not have been written by someone with working command of the subject.

IOP Publishing's Pattern of Retractions

IOP Publishing has now become the most visible case study in how a publisher responds to systematic tortured-phrase contamination. In February 2022, IOP retracted 350 papers from two conference proceedings, citing what it described as systematic manipulation of the publication process. Those retractions triggered changes in IOP's submission handling. The publisher added automated checks against the known tortured-phrase list maintained by Cabanac and colleagues, and announced it would apply retrospective screening to older proceedings content.

The October 2026 batch reflects that retrospective work continuing. The roughly 80 papers retracted that week came from three separate IOP conference series: Journal of Physics: Conference Series (29 papers), IOP Conference Series: Earth and Environmental Science (8 papers), and IOP Conference Series: Materials Science and Engineering (43 papers). The retraction notices cite the Problematic Paper Screener as the detection tool. IOP's running total for this specific problem, according to Cabanac as reported by Retraction Watch, now exceeds 1,000 papers.

That figure represents what one publisher found after systematic screening of its own proceedings catalog. It almost certainly underestimates the global scale. The Problematic Paper Screener's public database had flagged roughly 12,000 papers across all publishers as of late 2023, and the list of known phrases has continued to grow since then. IOP has been notably transparent about its retraction decisions. Many publishers have not been.

How the Problematic Paper Screener Works

The Problematic Paper Screener, built and maintained by Cabanac and collaborators at the Université Grenoble Alpes, runs weekly searches across approximately 130 million scholarly papers in the Dimensions.ai database. It uses a curated list of phrase pairs, each consisting of the tortured version and the original scientific term it misrepresents. When a paper contains five or more matches from this list, it is automatically flagged for human review.

The tool is not limited to tortured phrases. It includes eight detection modules, covering meaningless computer-generated text from SCIgen (a program designed to produce grammatically plausible but content-free papers), Mathgen (the mathematics equivalent), and as of 2025, patterns associated with unrevised large language model output. The phrase list itself is collaborative. Researchers who identify new instances can submit them, and the list is released under an open license. A March 2026 snapshot contained 7,942 phrase pairs.

The threshold of five matched phrases is a deliberate calibration. A single obscure phrasing choice in a non-native English paper would not trip the screen. Papers that reach the threshold have typically been processed through a spinner extensively, and the clustering of multiple matched phrases in a single document is a strong signal of systematic use rather than an isolated slip.

Why This Matters for Medical Journals Specifically

The most affected publisher categories in the Problematic Paper Screener's dataset are conference proceedings series and high-volume open access journals in engineering, materials science, and computer science. But the phrase list includes a substantial subset of biomedical terminology. "Bosom disease" and "bosom tumor" are on the list. So are spinner outputs for terms including "cardiovascular," "malignant," "clinical trial," and "polymerase chain reaction." Papers submitted to MDPI journals, a common destination for biomedical open-access publishing, appear in the screener's flagged database. Hindawi titles have appeared too.

This is not to suggest that mainstream clinical journals like the BMJ, JAMA, or NEJM are facing comparable contamination. Their editorial pipelines, submission volumes per editor, and pre-review screening are different in kind from what a low-cost conference proceedings series manages. The concern is the broader literature base that medical researchers cite. A systematic reviewer who finds a paper on PubMed, a preprint on bioRxiv, or a conference abstract cited in a secondary source has no reliable way to know whether that paper passed through spinner software before it was submitted. The Retraction Watch database holds thousands of retractions that list "concerns about text integrity" as a reason, a category that includes tortured phrases.

The downstream effect on systematic reviews and clinical guidelines has been documented separately. A September 2026 preprint showed papers from a known authorship-for-sale scheme, not specifically tortured-phrase papers but a related category of fraudulent work, cited in more than a dozen clinical guidelines. The mechanisms of contamination are different, but the failure mode is the same: fraudulent or low-quality papers enter the literature, get cited, and shape downstream recommendations before anyone notices.

Legitimate Authors and the Risk of False Positives

Honest researchers rarely use spinner tools. But there are two scenarios where a legitimate paper can trip a tortured-phrase screen, and both are worth understanding.

The first is machine translation without adequate expert review. A clinical researcher in a non-English-speaking country who uses a standard machine translation service to render their methods section into English, then submits without having a fluent expert check the result, may unknowingly produce output that contains some of the same phrase substitutions the spinner list is designed to catch. The phrases may be fewer and less systematic than those in a deliberately spun paper, but the detection system flags on phrase matches, not on intent. The safest practice for authors translating from another language is to have a fluent medical English editor review the entire manuscript before submission, specifically looking for any phrasing that a domain expert would find strange.

The second scenario is less common but worth noting. Some AI writing assistants, particularly lower-cost tools that are not specifically trained on scientific language, occasionally substitute general synonyms for technical terms when rephrasing text for clarity. This is the same behavior the spinner tools produce, and the output can end up in a paper whose authors did not intend any evasion. The disclosure problem and the detection problem overlap here: an author who used an AI writing tool and did not review every suggestion carefully may have introduced phrase substitutions they are not even aware of.

The Detection Arms Race and Where It Is Heading

Tortured-phrase detection is one of the oldest automated integrity tools, and it has already entered a second generation. The original approach depended on a fixed list of known phrase pairs. That list grows as researchers identify new examples, but the underlying spinner tools also evolve. As AI text generation becomes more widely accessible, the crude paraphrasing that produces obvious tortured phrases is being supplemented by more sophisticated output that does not rely on synonym substitution at all. A paper drafted entirely by a large language model, then lightly edited by a human, may not contain a single phrase from the Cabanac list.

Publishers are aware of this. The Problematic Paper Screener added an AI text fingerprint detector alongside its tortured-phrase module in 2025. COPE's updated retraction guidelines from August 2025 explicitly address AI-generated content as a ground for retraction when it was undisclosed. Detection tooling being developed at several publishers goes beyond phrase matching toward statistical analysis of vocabulary diversity, sentence-level perplexity, and clustering of author submission patterns. The transition is from looking for specific tells to looking for distributional anomalies.

What this means practically is that the category of automated fraud is not disappearing, but it is shifting. The papers most likely to be retracted in future years for AI or automated-writing problems are not the ones containing obvious spinner artifacts. They are the ones where a human and an AI collaborated in ways that were never disclosed and where the result contains subtle inaccuracies, hallucinated citations, or fabricated data that a more advanced check eventually catches.

The October 2026 IOP batch represents the tail end of one detection cycle, not a new problem. It is worth understanding because the techniques that generated those papers are still in use, the phrase list is still growing, and medical publishers are increasingly running their own versions of the same check. But it is also worth understanding as a template for what comes next: a slow-moving wave of retrospective audits catching content that should never have been published, years after it entered the literature and began accumulating citations.

What to Do If Your Manuscript Gets Flagged

If a journal contacts you about a tortured-phrase concern in a manuscript that you know to be genuinely your own work, the response process follows a similar path to other integrity inquiries. The key difference from, say, a concern about a figure is that you need to identify the source of the suspicious phrasing specifically. That almost always means tracing which draft the phrase appeared in, who wrote or edited that draft, and whether any automated tool was applied at that point.

If you can demonstrate that the phrase appeared because of a machine translation step that was not reviewed carefully, that explanation is credible and editors generally distinguish it from intentional fraud, though the resolution may still require a revision or correction to the manuscript text. If you cannot identify a source at all, that is a problem, because it suggests the draft may have passed through multiple hands or tools without adequate oversight.

In all cases, do not respond to the journal's inquiry by simply asserting that the manuscript is original. That assertion is what the phrase itself contradicts. Engage with the specific phrasing the journal has identified, explain how it could have appeared, and provide what documentation you have of the manuscript's development history. A version-controlled document record, correspondence with co-authors, or a record of the translation service used are all relevant.

Checking the Citations You Are Building On

Beyond the question of your own manuscript, there is a separate responsibility that falls on medical researchers building literature reviews and systematic searches. Papers that contain tortured phrases are not inherently wrong in their conclusions. Some were produced from real research conducted in good faith, then processed through spinner tools to evade plagiarism checks at submission. Whether the underlying data is reliable is a separate question from whether the text is original.

This means that a paper on the Problematic Paper Screener's flagged list is a paper to investigate, not necessarily to discard from a review. But it is a paper to investigate. The Retraction Watch database, the Crossref retraction API, and scite.ai all provide ways to check whether a paper you are considering citing has been retracted or flagged. For a systematic review, verifying the status of every included paper is part of the methods, and it belongs in the methods section as a discrete step.

The IOP retractions from October 2026 are now in the literature record. Any paper that cites one of those retracted works, whether in a biomedical context or not, inherits a small amount of the associated uncertainty. Screening your citation list at the time of final submission is the only way to know that you have not built on papers that were retracted in the same week your manuscript was under review.

Further Reading