|
The scale of fraudulent research in cancer science may be far larger than previously estimated. A team of researchers from Queensland University of Technology, the University of Sydney, and CNRS in France has applied a fine-tuned BERT language model to screen 2,647,471 cancer research papers published between 1999 and 2024. The result: 261,245 papers — 9.87 percent — showed writing patterns characteristic of paper mills, commercial operations that mass-produce fabricated or plagiarised manuscripts for sale to researchers.
The study, published in The BMJ, represents the largest systematic screen of cancer literature for paper-mill contamination. The rate has risen steeply over time, from roughly 1 percent in the early 2000s to a peak exceeding 16 percent in 2022, before dipping slightly in 2023-2024.
How the detection works
The model, built on Google’s BERT base uncased (110 million parameters), was fine-tuned on 2,202 papers that had been retracted from the Retraction Watch database with a “Paper Mill” tag. These served as positive examples — papers whose writing style, structure, and phrasing pattern matched the output of template-based manuscript factories. For negative controls, the team selected 2,202 papers from high-impact journals and countries with low paper-mill representation (Sweden, Finland, Norway, Taiwan).
The model reads titles and abstracts sentence by sentence, producing a probability score for each. The final classification is the mean of all sentence-level scores. In internal validation, the model achieved 91 percent accuracy with 87 percent sensitivity and 96 percent specificity. Against external validation sets — 3,094 papers independently flagged by image-integrity experts — accuracy rose to 93 percent.
Geographic and publisher patterns
The flagged papers are not evenly distributed. By first-author country, China accounted for 177,907 flagged papers — 36 percent of all cancer papers with Chinese first authors. Iran (20 percent), Saudi Arabia (16 percent), Egypt (15 percent), and Pakistan (13 percent) followed. The United States had 10,511 flagged papers, roughly 2 percent of its cancer output.
By publisher, the highest rates were concentrated in smaller operations: Verduci Editore (roughly 67 percent of its papers flagged, largely in the European Review for Medical and Pharmacological Sciences), International Scientific Literature (45 percent, largely in Medical Science Monitor), and Spandidos Publications (38 percent, or 19,043 papers). Among major publishers, Springer Nature, Elsevier, and Wiley each had roughly 10 percent of their cancer papers flagged, though in absolute numbers these were substantial — 40,293 for Springer Nature alone.
Gastric cancer research had the highest rate by cancer type at 22 percent, followed by bone/osteosarcoma (21 percent) and liver (20 percent). Fundamental cancer biology was more contaminated than clinical areas: survivorship, epidemiology, and health policy research had flagging rates below 2 percent.
Three journals already testing it
Three journals from a major publisher have already integrated the model into their online submission systems to screen cancer-related manuscripts in real time. The journal names have not been disclosed — the authors intentionally withheld them to prevent paper mills from adapting their templates.
“Its a statistical screen, not an attribution of misconduct,” the authors caution in the paper. Given an estimated true prevalence of roughly 10 percent in the literature, approximately 30 percent of flagged papers would be false positives.
Important caveats
The model was trained on Retraction Watch data that overrepresents Chinese-authored retractions, which may introduce geographic bias. Additionally, because the model is a deep neural network, the specific features it detects are not directly explainable. The authors acknowledge that the model may penalise formulaic English from non-native speakers, potentially conflating language patterns with fraud patterns.
There is also an arms-race problem: as detection tools improve, paper mills will adapt their templates. The rise of generative AI, the authors note, further blurs the boundary between genuine and fabricated text.
Sources
- Scancar B, Byrne JA, Causeur D, Barnett AG. “Machine learning based screening of potential paper mill publications in cancer research: methodological and cross sectional study.” The BMJ 392:e087581, 2026. DOI: 10.1136/bmj-2025-087581
- Queensland University of Technology press release via ScienceDaily