AI-Generated Papers and Journal Integrity

Two quite different things are discussed under one heading, and almost all the confusion comes from that. One is a researcher using a model to draft, edit or translate work they did. The other is fabricated content submitted to inflate a publication record. The first is a disclosure question. The second is fraud, and it is not new.

Two problems wearing one name

A non-native English speaker using a model to make their methods section readable has done nothing wrong and has improved the literature. A paper mill generating plausible manuscripts at volume has committed fraud, and would have done so with or without a language model — mills existed, using image manipulation, template text and fabricated data, long before this technology arrived.

Keeping them apart matters because they call for opposite responses. The first needs a disclosure norm and nothing else. The second needs content verification, and content verification does not care what tool produced the content. Any policy built around detecting machine text will punish the first group and miss most of the second, because fabricated research that has been lightly rewritten is indistinguishable from careful assisted writing.

It is also worth being clear about where the demand comes from, because it explains why no technical measure will resolve this. Paper mills exist because publication counts are used as a proxy for research contribution in hiring, promotion and institutional ranking, in systems large enough that buying an authorship is a rational purchase for some buyers. Generative tools lowered the cost of supplying that demand; they did not create it. A detector, even a perfect one, sits downstream of an incentive that would simply route around it — which is why the interventions with the best track record are the ones that attack verifiability, such as requiring data and code, rather than the ones that attack production.

What the artefacts look like

• Leftover interface text. Phrases that belong to a chat interface rather than to a paper — a preamble about being a language model, or a stray instruction to continue — have appeared verbatim in published articles. What this reveals is not that a model was used. It is that the author, the reviewers, the editor and the production process all failed to read the text, which is a far worse finding.
• Nonsense figures. Generated images with garbled labels and anatomically impossible content have been published and subsequently retracted. Same lesson: the figure was never examined.
• Tortured phrases. Standard terms replaced by odd synonyms — the sort of substitution a paraphrasing tool makes to evade plagiarism software. This artefact predates language models entirely and has been catalogued by integrity screening projects for years, which is a useful reminder that the underlying behaviour is older than the current tooling.
• Fabricated references. The most consequential of the four, because it survives copy-editing, looks completely normal, and corrupts the citation graph. See why plausible references get generated for the mechanism.
• Confident, specific, unsupported claims. The hardest to spot and the reason the others matter: the failure mode of a language model is fluency without grounding, which is precisely what peer review is least well equipped to catch quickly.

Why detection is the wrong instrument

Statistical detectors for machine-generated text have two problems that are not going to be engineered away, and a third that is arithmetic.

They are biased against non-native English writers. Published work evaluating detectors on essays by non-native English speakers found them flagged at high rates, for a straightforward reason: the features detectors key on — limited vocabulary variety, regular sentence structure, low unpredictability — describe careful second-language writing as well as they describe generated text. A tool that preferentially accuses one group of authors is not usable in an integrity process regardless of its average accuracy.

They are trivially defeated. Paraphrasing, editing or asking for a different style moves text out of the detected region, so the detector selects against exactly the people who did not try to hide anything.

And the base rate defeats what remains. Run any detector with a small false-positive rate across a large submission stream and the number of wrongly flagged honest authors will be large in absolute terms — potentially larger than the number of genuine cases, depending on how rare the genuine cases are. A flag is therefore not evidence about an individual manuscript, and using it as though it were is how careers get damaged over a probability score with no chain of custody.

Where policy has settled

There is now broad agreement among major publishers, editors and publication ethics bodies on a small set of points. They are stable enough to state, though this page carries a dated decay marker precisely because the details are still moving.

The checks that actually work

Every one of these targets whether the content is true rather than how it was produced, which is the right target and is also the only one that is defensible when challenged.

• Resolve every reference. Mechanical, cheap, and it catches the artefact that does the most damage. This alone would have stopped a large share of the visible cases.
• Check that cited work supports the claim. A real paper cited for something it does not say passes every automated check there is, and is the residue left after the easy checks.
• Ask for the data and the analysis code. Fabrication is much harder to sustain when the artefacts have to be consistent with each other.
• Read the figures. Most published generated-content embarrassments were visible to anyone who looked at the images.
• Treat any detector output as a prompt to look, never as a finding. If looking produces nothing verifiable, there is nothing to raise.