Open data sets and AI tools can be used to mass-produce low-quality, redundant papers.Credit: Tutatama/Alamy
An analysis of a literature database finds that text-generating artificial intelligence (AI) tools — including ChatGPT and Gemini — can be used to rewrite scientific papers and produce ‘copycat’ versions that are then passed off as new research.
In a preprint posted on medRxiv on 12 September1, researchers identified more than 400 such papers published in 112 journals over the past 4.5 years, and demonstrated that AI-generated biomedicine studies could evade publishers’ anti-plagiarism checks.
Low-quality papers based on public health data are flooding the scientific literature
The study’s authors warn that individuals and paper mills — companies that produce fake papers to order and sell authorships — might be exploiting publicly available health data sets and using large language models (LLMs) to mass-produce low-quality papers that lack scientific value.
“If left unaddressed, this AI-based approach can be applied to all sorts of open-access databases, generating far more papers than anyone can imagine,” says Csaba Szabó, a pharmacologist at the University of Fribourg in Switzerland, who was not involved in the work. “This could open up Pandora’s box [and] the literature may be flooded with synthetic papers.”
Redundant research
To investigate, researchers screened association studies — those that statistically link a variable to a health outcome — that were based on data from the US National Health and Nutrition Examination Survey (NHANES), a huge open repository of data on the health, diet and lifestyles of thousands of people.
They focused their search on studies they defined as ‘redundant’, meaning that the work tested the association between the same variable and health outcome as other research did, but analysed a subtly different subset of the actual data — including results from different survey years, for example, or participants of a different age or sex.
AI linked to explosion of low-quality biomedical research papers
Their search of the PubMed index of biomedical literature revealed 411 redundant studies published between January 2021 and July 2025. Most were simple ‘repeat’ cases, involving two publications that were almost identical. But three associations had a particularly high number of duplicate studies — six papers apiece — some of which were published in the same year.
This “shouldn’t be happening, and it doesn’t help the health of the scientific literature”, says co-author Matt Spick, a biomedical scientist at the University of Surrey in Guildford, UK.
Most publishers have checks in place to prevent researchers submitting the same research to multiple journals, but Spick and his colleagues suspect that AI tools are being used to evade these.
Avoiding detection
To test whether AI can help to produce multiple papers from the same data set, the researchers used OpenAI’s chatbot ChatGPT and Google’s Gemini to rewrite three of the most heavily redundant articles identified by their analysis (each reported a particular association that had already been published five or six times). The researchers prompted the LLMs to use the information in each paper, and NHANES data, to produce a new manuscript that could avoid plagiarism detectors.
“We were shocked that it worked straight away,” says Spick. “They weren’t perfect, and the LLMs did create some errors. It took two hours of cleaning-up work for each manuscript.”
Rise of ChatGPT and other tools raises major questions for research
When analysed with a plagiarism-detection tool used by many publishers, the AI-generated manuscripts did not score at a level that would be considered problematic by editors. This shows that LLMs “can produce something that is derivative of everything that’s gone before and doesn’t include anything new. But it’ll still pass plagiarism checks”, says Spick. This, in turn, makes it more difficult to distinguish between researchers who are producing genuine research using public-health data sets such as NHANES, and others who are deliberately creating redundant papers using LLMs, the authors note.
“These are completely new challenges for the editors and the publishers,” says Igor Rudan at the University of Edinburgh, UK, who studies global public health and is joint editor-in-chief of the Journal of Global Health, which introduced new guidelines for researchers submitting research on open-access data sets in July. “When we first tried the LLMs, we immediately realized that this would become an issue, and this preprint confirms it,” he adds.
Serious challenge
In July, Spick and his colleagues reported2 that there had been a surge in low-quality, formulaic publications using NHANES and other publicly available health data sets that they suspected had been fuelled by AI. The current analysis found a sharp rise in redundant NHANES studies after 2022 — the year ChatGPT was publicly released.
Some publishers, including Frontiers, based in Lausanne, Switzerland, and the open-access Public Library of Science (PLOS) in San Francisco, California, have tried to address this by tightening up their editorial rules for accepting studies that are based on open-access health data sets such as NHANES.
“AI-driven redundancy, in general, poses a serious and ongoing challenge to publishers,” says Elena Vicario, Frontiers’ head of research integrity. Frontiers published 32% of the 411 redundant papers identified in the preprint, with 132 articles appearing across 11 of its journals. Vicario says that these papers pre-date policies that were introduced by the publisher earlier this year, and would not be published if they were submitted today. Frontiers journals have rejected 1,382 NHANES-based submissions since the introduction of their policy in May.
The fight against fake-paper factories that churn out sham science
Springer Nature journals published 37% of the papers flagged by the preprint, including 51 articles that appeared in the journal Scientific Reports (Nature’s news team is editorially independent of its publisher). “We take our responsibility towards maintaining the validity of the scientific record very seriously and all papers referenced by this preprint will be investigated and action taken where appropriate,” says Richard White, editorial director for Scientific Reports, which, he says, has rejected more than 4,500 NHANES-based submissions since the start of 2024.
“There are differences in opinion regarding the value of some analyses utilizing NHANES and similar data sets, and we are committed to both supporting the whole community and ensuring what we publish adds value,” White adds. “Our focus is on getting the right checks in place to remove unethically produced or meaningless research, whilst still publishing papers using such data sets that represent valid and valuable additions to the scientific literature.”
“We are alive to concerns around inappropriate use of these data sets and have been taking sustained action.”
