The Committee on Publication Ethics (COPE) maintains the most widely adopted retraction guidance in scholarly publishing, and its retraction guidelines were revised in 2025 to name two grounds that earlier versions did not spell out: undisclosed use of artificial intelligence and third-party involvement such as paper-mill-supplied authorship. Neither is a new category of misconduct in practice — both have been driving real retractions for several years — but naming them explicitly gives editors a citable basis for retraction notices and closes an ambiguity that publishers had been resolving inconsistently case by case.
This piece explains what the current COPE retraction guidance actually says, how it differs from the guidance CASRAI has already covered elsewhere on this site, and what it changes operationally for editors and research-integrity offices.
What COPE retraction guidance is for
COPE guidelines are norms, not law. COPE has no enforcement power over any journal, publisher, or institution; it sets a shared standard that member publishers agree to apply, and its guidance documents function as the reference editors point to when justifying a decision. For the mechanics of how COPE guidance is structured and applied day to day, see CASRAI’s COPE Guidelines Explained and COPE Flowcharts Explained.
Retraction itself is a correction-of-the-record mechanism, not a punishment. This distinction matters because it is routinely misunderstood outside publishing and integrity circles: a retraction removes a paper’s standing as part of the citable literature, it does not erase the record, and it is not equivalent to a disciplinary finding against an author. See CASRAI’s retraction definition and, for the common confusion with an unrelated concept, retraction vs. redaction.
The established grounds, and what’s newly named
COPE’s long-standing retraction grounds cover findings that are unreliable due to error or fabrication, prior publication without proper attribution or justification (redundant/duplicate publication), plagiarism, unethical research (for example, without proper ethics approval), compromised or manipulated peer review, and undisclosed conflicts of interest. These remain the backbone of the guidance.
What the 2025 revision adds is explicit naming of two grounds that previously had to be reasoned into the existing categories rather than cited directly:
- Undisclosed AI-generated content. Where a paper’s text, data, images, or analysis involved undisclosed AI generation in a way that undermines confidence in authorship, originality, or the reported results, that non-disclosure is now a named basis for retraction.
- Third-party and paper-mill involvement. Where authorship cannot be verified, accountability for the work is uncertain, or a paper shows signs of systematic, coordinated manipulation — fabricated or purchased authorship slots, fake or compromised peer review tied to a paper mill, citation manipulation rings — that is now named directly rather than treated only under the broader “unreliable data” or “compromised peer review” headings.
Guidance current as of this writing also recommends that retraction notices acknowledge third parties who raised the concern (such as sleuths or data-integrity reviewers) when relevant and permitted, and it emphasizes timeliness and consistent notice content as standing expectations rather than optional practice.
Why non-disclosure, not AI use itself, is the failure
It is worth being precise about what actually triggers retraction under this ground. Most major publishers’ author policies permit disclosed, appropriately limited AI assistance in manuscript preparation (for example, language editing) provided it is declared and the authors retain full responsibility for accuracy. The retraction trigger is not “AI was used” — it is that use was concealed, and the concealment itself misrepresents how the work was produced. This mirrors how COPE treats other disclosure failures, such as undeclared conflicts of interest or ghost authorship: the underlying act is often not disqualifying on its own, but hiding it breaks the trust the publication record depends on. CASRAI covers the compliance side of this gap in the AI disclosure gap between policy and compliance, and the related question of undisclosed AI specifically flagged in manuscripts is addressed in COPE’s own case guidance on suspected AI-generated manuscripts.
Third-party authorship and paper mills
Paper mills sell authorship slots, fabricate data, and in some cases compromise peer review by supplying reviewer identities the mill itself controls. Because the resulting papers are produced at scale and follow reusable templates, they tend to surface as clusters of retractions rather than isolated cases — a pattern CASRAI has tracked across several publisher actions, including the Wolters Kluwer transplant-journal retractions, the ASTM journal retractions, and the Frontiers peer-review manipulation network. Naming “authorship cannot be verified” and “accountability is uncertain” as retraction grounds in their own right gives editors language for exactly this pattern, rather than requiring them to force it under a data-fabrication or peer-review-manipulation heading that may not fit the specifics. For the coordinated response at the publisher and industry level, see CASRAI’s coverage of the United2Act paper mill coalition and COPE’s relaunched paper mills working group. For definitions, see paper mill and COPE (Committee on Publication Ethics).
Retraction vs. correction vs. expression of concern
The decision logic COPE guidance sets out is straightforward in principle, harder in practice:
- Correction is appropriate when a discrete error does not undermine the paper’s overall reliability, conclusions, or scholarly integrity — a mislabeled figure, an author-affiliation error, a citation fix.
- Retraction is appropriate when the reliability of the findings, the validity of authorship, or the integrity of the underlying research process is compromised such that the work should no longer stand as part of the literature.
- Expression of concern (EoC) is used when credible evidence of a problem exists but cannot yet be resolved — typically because an institutional investigation is ongoing and the publisher does not have enough information to retract or clear the paper. An EoC signals readers to treat the paper with caution while that process plays out.
See CASRAI’s expression of concern definition and the practical walkthrough in how a retraction actually happens and how long does a retraction take.
Process obligations: what a compliant retraction notice does
Regardless of the ground cited, COPE guidance sets consistent process expectations for how a retraction is executed, not just when one is warranted:
- The retraction notice states who is retracting the article (authors, editor, publisher, or a combination) and why, in terms specific enough to be meaningful to readers.
- The original article remains accessible — it is not deleted — and is clearly and persistently marked as retracted on every version (HTML, PDF, and any indexed copy).
- The retraction is linked bidirectionally to the original article, and the article’s DOI persists and continues to resolve to the (now-marked) original plus the retraction notice.
- Retraction metadata propagates to the indexes and services that track it — Crossref’s retraction metadata and the Retraction Watch Database being the two most consulted downstream sources — so the retracted status follows the paper wherever it is subsequently cited or discovered.
For notice-drafting specifics, see CASRAI’s how to write a retraction statement and the retraction statement definition.
What this means operationally
For research-integrity officers and institutional research offices, the practical effect is a clearer citation point when a case involves undisclosed AI use or fabricated/purchased authorship — editors no longer need to stretch an existing category to justify retracting on those grounds. But it does not resolve the longstanding coordination problem between publisher and institutional processes: publishers generally will not wait indefinitely for an institutional misconduct investigation to conclude before acting (guidance explicitly discourages open-ended delay pending an institution or an uncooperative author), while institutions often need the publisher’s findings, or at least its timeline, to scope their own review. That mismatch in timelines predates this revision and continues to be the main friction point institutions report when a case involves both a publisher-side retraction process and a parallel institutional misconduct inquiry.
For editors, the operational change is smaller but real: retraction notices citing AI non-disclosure or unverifiable/third-party authorship as the ground now have direct guidance language to draw on, which should improve consistency in notice wording across publishers — one of the persistent quality gaps Retraction Watch and others have documented in retraction notices generally.
Frequently asked questions
Does this mean using AI to help write a paper is now grounds for retraction?
No. The ground is undisclosed AI-generated content, not AI assistance itself. Most publisher policies permit disclosed, limited AI assistance (such as language editing) with authors retaining responsibility for accuracy. Concealing that assistance, or using AI to generate results or content presented as the authors’ own undisclosed work, is what triggers the retraction ground.
Is a paper mill retraction different from a fabricated-data retraction?
They can overlap, but the grounds are conceptually distinct. A paper-mill-linked retraction is grounded in unverifiable authorship, uncertain accountability, or systematic manipulation of the publication process (including peer review); fabricated data is a separate, older ground concerning the reliability of the findings themselves. A single paper can be retracted on both grounds at once.
Why use an expression of concern instead of retracting immediately?
When there is credible evidence of a problem but the investigation needed to confirm it — often an institutional inquiry — is still underway, an EoC lets the publisher flag the concern to readers without prejudging an outcome the publisher cannot yet confirm on its own.
Does a retracted article get removed from the journal’s website?
No. COPE guidance requires the original article to remain accessible, clearly marked as retracted, with its DOI intact. Removal is reserved for narrow legal exceptions (for example, defamation or copyright issues), not for standard retraction.
