A new study published in The BMJ used a BERT-based 'scientific spam filter' to screen 2.6 million cancer studies published between 1999 and 2024. The AI model, trained on 2,202 retracted papers from known paper mills, flagged 261,245 papers — 9.87% of the total — as suspicious. The findings suggest that fraudulent research may be far more common in the cancer literature than previously assumed.
How the AI filter works
The tool is a type of language model called BERT, which learns patterns in text. Researchers trained it on a set of papers that had been retracted because they were produced by paper mills — operations that churn out fake or plagiarized studies for profit. The model then scanned the full text of 2.6 million cancer papers, looking for similar writing patterns. It flagged any paper whose language resembled the retracted mill output.
Adrian Barnett, a biostatistician at Queensland University of Technology, led the study. He and his team did not manually verify each flagged paper, so the 9.87% figure is an upper bound. Some flagged papers may be legitimate but happen to share stylistic quirks with mill papers. Still, the scale is striking.
A growing problem over time
The proportion of flagged papers rose sharply across the study period. In the early years, around 2000, only about 1% of cancer studies were flagged. By the end of the period, that share had climbed to nearly 10%. The trend suggests paper mills have become more active — or more sophisticated at evading traditional checks.
Cancer research is a high-stakes field. Flawed studies can mislead other scientists, waste grant money, and even affect patient care if results are used to guide treatment decisions. The study does not name specific journals or authors, but it points to a systemic vulnerability in the publishing system.
What the findings don't tell us
The AI filter is a screening tool, not a final verdict. It cannot prove a paper is fraudulent, only that its writing looks suspicious. The authors note that some legitimate papers might use similar phrasing — for example, if they describe standard methods in a formulaic way. But the sharp increase over time is hard to explain away as coincidence.
Paper mills are a known problem in academic publishing. They sell authorship or produce entire manuscripts for researchers who need publications. Retraction databases have grown, but catching mill papers after publication is slow. A tool like this could help journals spot problems before they print.
Next steps for the research community
The study is a proof of concept. Barnett and his team have not released the filter for general use, but they argue that journals and funders should consider adopting similar AI screening. The next question is whether publishers will act. Some already use plagiarism checkers, but pattern-based detection is newer. The BMJ study gives them a reason to look harder.
