AI Flags a Quarter Million Suspicious Cancer Research Papers

A team at Queensland University of Technology (QUT) in Australia built a tool based on the BERT language model that scanned 2.6 million cancer research papers published between 1999 and 2024. The program flagged more than 250,000 of them as suspicious, showing the same linguistic patterns as articles previously retracted for ties to so-called paper mills, companies that produce fake research to order.

The study was led by Professor Adrian Barnett of the School of Public Health and Social Work and the Australian Centre for Health Services Innovation at QUT, working with an international group of researchers. The results were published in The BMJ, though the study only gained wide attention in mid-July 2026, when press releases from the university and science agencies spread through global media.

How the filter works

Rather than analyzing images, charts, or raw data, access to which is practically impossible across millions of publications, the researchers focused solely on titles and abstracts. The BERT model was trained on more than two thousand oncology papers already retracted for proven ties to paper mills, teaching it to recognize their characteristic patterns: repetitive phrasing, rigid sentence structures, and formulaic descriptions of methodology.

We've essentially built a scientific spam filter - Adrian Barnett, Queensland University of Technology

In tests on a set of verified examples, the tool correctly identified suspicious papers in 91 percent of cases within the training set and 93 percent on external data, with specificity exceeding 96 percent. That means the model rarely falsely flags legitimate research, though the authors stress that a flag is not proof of misconduct, only a signal for further review.

The scale of the problem

The most troubling finding is the trend over time. In the early 2000s, suspicious papers made up about 1 percent of annual oncology research output. By 2022 that share had climbed to 16 percent, averaging nearly 10 percent across the whole study period. The problem isn't confined to low-tier journals: the share of suspicious papers also rose in high-impact-factor titles, exceeding 10 percent in recent years.

The researchers also found clear geographic differences. Among papers linked to Chinese institutions, the flagged share reached 36 percent, well above the global average. Suspicious papers cluster most heavily in tumor molecular biology and early-stage lab research, and among cancer types, they are especially common in studies on stomach, liver, bone, and lung cancer.

Paper mills are companies that sell fake or low-quality research. They produce studies on an industrial scale - Adrian Barnett, Queensland University of Technology

Why this matters for patients

Oncology wasn't a random choice for this experiment. The findings of cancer research papers directly shape the design of clinical trials, funding decisions for therapies, and bedside medical practice. If the knowledge base underpinning further studies and meta-analyses contains tens of thousands of fabricated results, the risk of flawed clinical conclusions grows in ways that are hard to quantify, since fake data can be cited for years before anyone challenges it.

The paper mill problem has been building for years, but until now there was no tool capable of assessing its scale across millions of publications at once. Manual verification of individual papers by reviewers or research integrity investigators takes weeks or months, while paper mills churn out new manuscripts at an industrial pace, selling authorship slots and ready-made texts with substituted data.

What's next for the tool

Three scientific journals are already testing the system as part of initial editorial screening, before manuscripts are sent to outside reviewers. The QUT team stresses, however, that the model's output is not final proof of misconduct, and every flag must be reviewed by human research integrity specialists before any accusation is made against authors.

For Polish research institutions and journal publishers, this matters regardless of where the study was conducted. If BERT-based tools become part of the standard editorial process at global publishers, Polish researchers publishing in international oncology journals could face an additional automated text-screening step before their work reaches reviewers. It's a shift comparable to the introduction of plagiarism detectors a decade ago, except this time the goal is catching entire fabricated studies, not just copied passages.