Machine Learning Screen Flags Nearly 10% of Cancer Literature as Possible Paper Mill Output

A machine learning screen published in The BMJ has put a number on a problem that research integrity watchdogs have been warning about for years: roughly one in ten cancer research papers may have originated from so-called paper mills, the commercial operations that manufacture and sell fake or fabricated studies.

<cite index="26-1">The study was led by Baptiste Scancar, Jennifer A. Byrne, David Causeur, and Adrian Barnett, drawn from institutions including L'Institut Agro in Rennes, the University of Sydney, and Queensland University of Technology.</cite> <cite index="22-4">Barnett's team analyzed 2.6 million cancer studies published between 1999 and 2024.</cite> The tool they built doesn't look for duplicated images or fabricated gel bands. It works on text.

<cite index="17-2">The researchers used a machine learning approach known as BERT (bidirectional encoder representations from transformers) to analyze text from publication titles and abstracts.</cite> The model was trained to distinguish known paper mill output from genuine research, then turned loose on the full corpus. <cite index="26-2,26-3">The model achieved an accuracy of 0.91, and when applied to the cancer literature it flagged 261,245 of 2,647,471 papers -- 9.87% (95% CI 9.83 to 9.90) -- revealing a large increase in flagged papers from 1999 to 2024, including in the top 10% of journals by impact factor.</cite>

That last detail matters. The contamination isn't confined to predatory journals or low-prestige outlets. High-impact venues are affected too.

<cite index="26-4,26-5,26-6">More than 170,000 papers affiliated with Chinese institutions were flagged, accounting for 36% of Chinese cancer research articles, and most publishers had published substantial numbers of flagged papers. Flagged papers were overrepresented in fundamental research and in gastric, bone, and liver cancer.</cite>

<cite index="12-11">The rate of flagged papers rose from roughly 1% in the early 2000s to 16% by 2022</cite> -- a trajectory that tracks the documented expansion of paper mill operations globally. A March 2026 study in PNAS, cited separately by analysts covering this space, described paper mills as criminal organizations whose output is doubling every 1.5 years.

The stakes aren't just bibliometric. As Barnett told the QUT press office, <cite index="22-9,22-10">"Cancer research influences clinical trials, drug development and patient care. If fabricated studies make their way into the evidence base, they can mislead real scientists and ultimately slow progress for patients."</cite>

A few methodological cautions are worth naming. <cite index="23-1">The model achieved 91% internal accuracy and 93% external accuracy, with specificity above 96%.</cite> Those are solid numbers for a screening tool, but specificity and sensitivity aren't the same thing. A 91% accurate classifier applied to 2.6 million papers still produces meaningful rates of both false positives and false negatives. The authors are explicit that the tool is designed for screening, not adjudication -- a flagged paper isn't a confirmed fake, it's a paper that warrants closer scrutiny.

<cite index="25-14">The team found that nearly ten percent of the cancer research literature screened with the tool could have originated in paper mills, a percentage that exceeds the estimated prevalence of paper mill papers in biomedical research more broadly, indicating that cancer research is a major target of these fraudulent operations.</cite>

<cite index="9-12">As generative AI makes it easier to produce fraudulent papers, researchers are turning to AI-powered detection methods in response</cite> -- a dynamic that has the feel of an arms race with no obvious resolution. The tool the QUT team built is open and intended for use by publishers. Whether journals adopt it at scale, and what happens when paper mills learn to evade BERT-based screens, are questions the paper doesn't resolve.

The cross-sectional study was published in The BMJ (DOI: 10.1136/bmj-2025-087581). The underlying model and screening results were released publicly by the authors.

Sources cited:
- The BMJ (via PMC) (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12853418/)
- Queensland University of Technology News (https://www.qut.edu.au/news?id=203173)
- The Scientist (https://www.the-scientist.com/nearly-ten-percent-of-cancer-papers-flagged-as-potentially-fake-74185)
- KFF Health Information and Trust Monitor (https://www.kff.org/health-information-trust/how-ai-can-both-detect-and-enable-fraudulent-research/)
- ScienceDaily (https://www.sciencedaily.com/releases/2026/07/260714225538.htm)