AI Flags Over 250,000 Suspicious Cancer Research Papers in Landmark Scientific Integrity Study

A new machine learning system has flagged more than 250,000 cancer research papers published between 1999 and 2024 as potential products of paper mills — businesses that manufacture and sell fraudulent scientific manuscripts. The study, published in The BMJ on July 16, analyzed 2.6 million papers in total. What it found is less a smoking gun than a flare fired into a building where the exits are unclear.

What the AI actually found

Researchers at Queensland University of Technology trained a machine learning system using BERT, a language-processing model, to identify writing patterns common in papers previously linked to paper mills. When applied to 2.6 million published cancer studies, it flagged approximately 250,000 — just under 10% of the entire corpus.

The proportion did not hold steady over time. In the early 2000s, flagged papers represented around 1% of annual cancer research output. By 2022, that figure had climbed past 16%. The increase appeared across thousands of journals, with molecular cancer biology and laboratory-based research showing the highest concentrations. High-impact journals were not immune — the share of flagged papers in those publications has also risen, crossing 10% in recent years.

Being flagged is not a finding of misconduct. The system surfaces patterns worth examining — linguistic fingerprints shared with known paper mill publications — but every flagged paper still requires expert review before any conclusion can be drawn. The researchers are explicit about this. The AI raises questions. It does not answer them.

Why this matters beyond the lab

Cancer research that enters the scientific record does not stay there in isolation. Other researchers cite it, build on it, and sometimes use it to design clinical trials or treatment protocols. Fraudulent papers, if undetected, can misdirect funding, inflate treatment expectations, and quietly contaminate the evidence base that doctors use to make decisions. Earlier this year, a separate analysis in Nature found that cancer papers suspected of originating from paper mills were attracting significantly more citations than legitimate studies. Bad research, it turns out, spreads fast.

The problem is also getting harder to contain. A Northwestern University study published in 2025 found that suspected paper mill output is doubling every 1.5 years, roughly ten times faster than the rate of growth in legitimate scientific output. Retractions are not keeping pace. Of more than 32,000 fraudulent articles identified in that study, only 29% had been retracted.

Generative AI is making it worse

Generative AI tools are making it cheaper and faster to produce plausible-sounding scientific text. A Lancet audit published in May 2026, covering nearly 2.5 million biomedical papers, found that the rate of fabricated references climbed from about 1 in 2,828 papers in 2023 to roughly 1 in 277 by early 2026. The same tools now used to detect fraud are also being used to produce it, faster, and at a scale no editorial team can manually screen for.

Journals have started responding. The International Committee of Medical Journal Editors updated its guidelines in January 2026 to require disclosure of substantive AI use and to prohibit listing AI systems as authors. Wiley, Elsevier, and Springer Nature have all updated their policies. Whether those rules slow the trend is a separate question from whether they are the right rules.

What comes next

The BMJ study team has made their model available to journals and publishers for use in screening new submissions. The goal is not to automate rejection but to flag papers for closer human review before they enter the record, rather than years afterward. The harder work — reviewing the 250,000 already-published papers that were flagged — falls to journals, editors, and the wider scientific community. That will not happen quickly. But the AI has at least given them somewhere to start, and in a field where the consequences of fraud reach patients, that is not a small thing.