An AI Trained to Spot Fake Science Flags 250,000 Cancer Papers as Possible 'Paper Mill' Products

A new artificial-intelligence tool has exposed what its creators call one of the biggest integrity problems in modern science, flagging more than 250,000 cancer research papers that show the linguistic hallmarks of fraudulent “paper mills.”

The study, published in The BMJ, examined 2.6 million cancer papers released between 1999 and 2024. Led by Professor Adrian Barnett of Queensland University of Technology, working with an international team, the analysis found that a large slice of the literature contained writing patterns resembling those in studies already retracted over suspected fabrication.

“Paper mills are companies that sell fake or low-quality scientific studies,” Barnett said. “They are producing 'research' on an industrial scale, and our findings suggest the problem in cancer research is far larger than most people realized.” Such operations sell authorship slots and even complete, ready-made manuscripts, often padded with reused text, awkward phrasing and fabricated data or images.

To detect them, the researchers trained a language model called BERT to recognize the subtle textual “fingerprints” that repeatedly surface in known paper-mill products. “Most likely, they're relying on boilerplate templates which can be detected by large language models that analyze patterns in texts,” Barnett said. In validation tests, the tool correctly flagged problematic papers with roughly 90% accuracy.

Perhaps the most alarming finding is the trajectory. The proportion of potentially problematic cancer studies has climbed steadily over two decades — from around 1% in the early 2000s to more than 16% by 2022 — suggesting the fraud is accelerating rather than fading. A polluted literature carries real stakes, because clinicians and researchers build future treatments and trials on the foundation of published results.

The authors were careful to stress the limits of their approach. The AI is designed to raise questions, not deliver verdicts: every flagged paper still requires expert review before any conclusion about misconduct can be drawn. Even so, the researchers argue the tool offers journals, publishers and institutions a way to triage a flood of submissions that has overwhelmed traditional peer review — an early-warning system for a problem that has quietly grown into a crisis of trust.

The stakes extend well beyond academic bookkeeping. Cancer research guides clinical trials and treatment decisions, and a literature seeded with fabricated results can send scientists chasing false leads and waste scarce funding. Barnett and his colleagues argue that journals, publishers and research institutions need faster, automated ways to screen the flood of submissions that has swamped traditional peer review, using tools like theirs to prioritize which papers deserve a closer human look. They caution that the same large language models powering the detector can also be used to churn out ever more convincing fakes, making the contest between fraud and detection an escalating arms race that scientific publishing cannot afford to ignore.