At a glance
Global publishers are now expanding AI-based manuscript screening to detect paper mills and ghostwriting before peer review. See what labs should prepare.
Global publishers are moving paper mill detection out of the post-publication queue and into the submission system. The targets are an old problem with a new signature: coordinated groups selling authorship on manufactured manuscripts, and undisclosed writers shaping manuscripts that named authors sign. The instruments are artificial intelligence (AI)-based manuscript screening tools that check images, text, author metadata, and submission behaviour before editors send work to reviewers.
A paper mill is an organisation that produces fabricated or manipulated manuscripts at scale and sells authorship slots. Ghostwriting, in the stricter editorial sense, is an undisclosed writer making a substantial contribution to a manuscript while the named authors take credit. Both flood the same part of the pipeline: the point where a hurried editor first decides whether a submission deserves serious review.
A colleague, Dr. K, edits a materials science journal. On a Tuesday this spring, eleven manuscripts arrived from different first authors, different institutions, four countries, and two shared equipment signatures. Six carried the same instrument readout backgrounds and the same paragraph transitions. She flagged those six. The other five survived two reviewers each and a month of back-and-forth before falling apart over missing raw data. Multiply her Tuesday by every editor handling more than a dozen weekly submissions, and the reason publishers are reordering their screening becomes obvious.
What the screening tools are actually looking at
Paper mill detection is not a single test. It combines signals from several places, which is why publishers describe it as a suite rather than a filter. Image integrity checks look for duplications, rotations, splicing, and reuse across unrelated papers. Text checks flag repeated boilerplate and phrase echoes across unrelated submissions; they also flag writing patterns that do not match an author's prior published style. Metadata checks look for simultaneous submission, shared email accounts, shared IP addresses, and author networks that cut across too many journals.
The infrastructure works at two levels. At the platform level, STM, the International Association of Scientific, Technical and Medical Publishers, runs shared screening infrastructure. Publishers then add internal tests. The split matters because a journal can catch duplicate images without knowing what every other journal is screening. It can also catch a manuscript submitted to two places at once, which a single publisher's internal system would likely miss. The purpose is to shorten the time between a paper mill's first submission and the moment editors see a pattern.
Screening is now being run before peer review, not as an after-the-publication safety net. That is the sharpest change from five years ago, when most detection happened after a paper was live and often only after a reader spotted a suspicious figure on PubPeer.
Why ghostwriting sits next to paper mills in the queue
Ghostwriting in research publishing is not the same as an author using AI to smooth phrasing. It means a hired or undisclosed writer contributes substantially to the work and disappears from the byline. The International Committee of Medical Journal Editors (ICMJE) requires disclosure of writing assistance, and most journals now extend the requirement to AI-assisted drafting. Screening flags prose that looks operationally uniform across authors, including characteristic phrasing patterns, not because uniformity proves ghostwriting, but because it prompts an editor to ask better questions.
The tools are being built at the platform level and the publisher level. STM operates the STM Integrity Hub, which lets participating publishers check submissions against a shared corpus. The hub began as a duplicate-submission pilot and has expanded to include paper mill detection, reference ring patterns, image duplication checks, and signals intended to catch simultaneous submissions. Publisher-specific systems layer on top. Springer Nature has been testing Geppetto, which flags AI-generated or AI-assisted text, and SnappShot, which checks image integrity. Geppetto looks at the prose; SnappShot looks at figures a ghostwriter may have borrowed or edited from previous papers.
COPE, the Committee on Publication Ethics, has been clear about why the expansion is happening now. Its paper mill guidance tells editors to treat paper mill activity as a systemic problem, not a series of one-off frauds. The same guidance underpins United2Act, which COPE and STM run together. Our earlier piece on the COPE paper mill detection guidance covers the specific editorial steps.
What the retraction record tells us
Wiley's Hindawi portfolio is the clearest recent example of the scale. After discovering large-scale paper mill activity, Wiley retracted thousands of papers and later said it would wind down the Hindawi brand. The sequence is covered in our earlier reporting on the Wiley Hindawi mass retractions. Retraction Watch's public database, which we covered when it passed 50,000 retractions, remains the public record of how much damage was done before submission-level screening became common.
The shift to pre-publication screening is partly a response to those numbers. Editors used to find paper mill papers after publication, usually through image sleuths on PubPeer or post-hoc editorial audits. The corrections happened in public, with the journal's reputation as a co-author. Now publishers want to intercept the paper before peer review, which is cheaper and less publicly damaging. Retraction Watch still logs the cases that slip through.
False positives and the human reviewer problem
AI-based manuscript screening has a false-positive tail. Shared phrasing is common because introductions and methods sections are highly conventional. Researchers in a narrow subfield use the same phrases for the same protocols. Non-native English speakers often use AI writing assistants to polish text; that is not ghostwriting if disclosed, but it can trip Geppetto-style detectors. Editors are therefore told to treat flags as prompts for questions, not verdicts.
At Dr. K's journal, roughly one in nine submissions now triggers an image or text flag. About two-thirds clear the query when authors provide raw data and a contribution statement. That is the base rate worth remembering: a flag is a request for documentation, not a verdict. A legitimate immunology author from a small hospital had his first AI-polished submission flagged for a style mismatch. The editor asked for raw data and author contributions. He produced both, and the manuscript moved forward. The screening cost was a four-day delay, not a rejection.
- Keep raw images and original instrument files, not just cropped PowerPoint figure panels.
- Record any use of AI-based editing or translation tools in the cover letter and the author contributions section.
- Retain the version history of the manuscript so you can show who wrote which sections.
- If you are asked about a flag, respond with the raw file and the contribution statement first, not with a legal threat.
What this means for your lab
The practical risk is not that a real lab will be ejected for a false positive. The risk is that the lab will not be able to answer the editor's question quickly. Data storage is the difference. A lab that keeps raw blots, microscope files, manuscript drafts, and contribution statements in a shared folder can clear an image-integrity query the same day. A lab that relies on a departed postdoc's laptop cannot.
Journals are also asking more specific authorship questions. If a manuscript has a named statistician who did not see the final draft, or a medical writer who drafted sections but is not listed, that is now a disclosure requirement. Screening is expanding because author and content checks have become cheap enough to run on every submission, not just on suspicious ones.
The next concrete step for authors is to run a pre-submission check on your own figures and text. Find out what your target journal uses, read its AI and authorship disclosure form before you submit, and prepare the raw data story in advance. The screening system is looking for the pattern a paper mill leaves when nobody in the lab can produce the underlying files.
Photo by Arnaud Jaegers on Unsplash
- research-integrity
- paper-mills
- springer-nature
- stm-integrity-hub
- ai-manuscript-screening

