Manuscripts about manuscripts featuring citations about citations – paper mills sell scientific articles that are fictitious and created either by people or generative AI. | Photo: Philotheus Nisch / Connected Archives

When the journal Scientific Reports published an article on autism in November 2025, it was accompanied by an infographic that used Dadaist vocabulary (e.g., ‘Medical frymbial’, ‘drop-out totalbottl’), that featured human legs poking through solid objects, and that included a picture of a bicycle marked ‘score 0.93’.  It’s actually obvious that the infographic was generated by AI. But no one seems to have noticed before it was published.

This particular sham was quickly spotted, but sadly it’s just the tip of the iceberg. Many bogus works go unnoticed, because they are harder to recognise. They mostly come from research paper mills: companies that sell not only manuscripts – made either by hand from scratch or by employing generative AI – but also the possibility of becoming a co-author of a publication. Their clientele seek to inflate their lists of publications and their numbers of citations, which are key statistics in a scientific career.

“The first cases of large-scale fraud appeared about ten years ago”, says Anna Abalkina of the Free University of Berlin, who studies corruption in the academic world. Paper mills can be found in India, Pakistan and China, as well as in Russia and Latvia. Their clientele live all over the world, including in the West”.

There are several points in the publication chain where they can inject their output, especially when journals plan to publish special issues on a given theme and invite academics to act as guest editors. In some cases, the editors have then called on acquaintances to submit articles, turned a blind eye to quality, and used their networks to ensure positive peer reviews. When a group of people cite each other like this, it creates a citation cartel.

The AI ‘arms race’

One person tracing these networks is Thomas Stoeger, who wrote his PhD in biology at the University of Zurich before joining Northwestern University near Chicago. He has identified dozens of suspicious publishers by way of abnormally high numbers of articles published over a short period and by their large numbers of accepted manuscripts either submitted by other members of the network or subsequently retracted.

His study published in PNAS reports 32,000 articles suspected of being papermilled content. They represent only 0.01 percent of the 270 million scientific texts identified by the Open Alex catalogue, but it’s the rapid growth that worries Stoeger: “Their number has doubled every 18 months and far exceeds the number of articles retracted or reported on sites such as PubPeer or Retraction Watch. If this trend continues – in particular due to the use of AI in producing manuscripts – we could end up with hundreds of thousands of fraudulent articles published each year, and three-quarters of them may never be retracted”.

It’s scientists who identify many problematic articles, either once they’re published or when they’re posted as preprints. This is quality assurance work carried out free of charge and in their spare time. There are other, more discreet efforts being made upstream in the publishing houses, however. They say they have taken stock of the problem. At Springer Nature, for example, each manuscript submitted is screened multiple times before reaching a responsible publisher , says Chris Graf, the director of research integrity.

These tools detect the use of large language models, plagiarism and suspicious paraphrasing (e.g., ‘unrefined information’ for ‘raw data’), hidden prompts used to trick any AI-detectors, manipulated images or off-topic references. What this constitutes is an arms race between those committing fraud and those detecting it. Details are kept confidential “so as not to inform bad actors ”.

Frontiers is a Lausanne-based publisher specialising in open access and employing 66 people in its research integrity unit. The unit rejects around 40,000 manuscripts each year, representing more than a quarter of total submissions and half of the refusals.

“Until recently, publishing houses worked in silos”, says Elena Vicario, the unit’s director. “But paper mills represent a global problem that requires collaboration among publishers”. Using the STM Integrity Hub platform, they share information on manuscripts suspected of being fraudulent, for example, multiple parallel submissions or suspicious citation numbers.

The fatal weapon: de-indexing

Publishers have also stepped up retraction processes. Some blatant cases of AI-generated content have been removed after only a few weeks, or even a few days. But Abalkina sees these efforts critically. “These instances of quick reactions relate to publicised cases and are not often seen in other, lower-profile cases. I think commercial publishers have little incentive to really fight paper mills”.

The least reliable journals potentially face a heavy sanction: de-indexation. Every year, major bibliographic catalogues, e.g., Scopus and Web of Science, remove a hundred journals from their corpus for not having respected quality standards, especially in relation to peer reviews and the number of fraudulent articles, according to Stoeger’s study. Scientists therefore become less interested in submitting an article to them, because it may not lead to any career benefits. In one case, the de-indexing in 2023 and 2024 of several journals published by the Basel-based giant MDPI led to a marked decline in the number of manuscripts submitted to it, according to Abalkina.

The reluctant author

There’s no telling where fraud leads, as the disinformation specialist Federico Germani of the University of Zurich testifies. In November 2025, he was browsing the open-research platform ResearchGate, only to discover his listing as a co-author of a manuscript he’d never seen before. “I tried to contact the authors, but found no trace of them. They are clearly false identities. But why were they created? One possibility is that paper mills want to attract future customers by showing off their capabilities and add the name of scientists affiliated with a well-known university to boost their credibility”.

Another reason for creating virtual identities is to act as an editor or peer-reviewer without assuming any risk, says Abalkina. The identities are then used to facilitate the publication of fraudulent articles or to abuse the system by citing articles of scientists who have paid for the service. It’s possible to buy citations in bulk on the internet. In one example, it cost a thousand dollars for a hundred citations. This is the same way that click farms are used to gain followers on social networks and to manipulate customer reviews on e-commerce sites.

These growing quantities of fraudulent science have multiple consequences. They overburden the publication process, which is already under pressure, and they bias the evaluation of academic careers. Scientific literature therefore becomes polluted, especially synthesis work such as meta-analyses and systematic reviews, leading to the risk of erroneous information appearing in search-engine results and AI-training datasets.

A central difficulty lies in the gap between good research practices and legislation, says Abalkina. “Purchasing pre-written articles, citations or co-authorship blatantly violates scientific ethics and integrity. But in many countries, these practices are not illegal”.