Researchers built a BERT-based text classifier using 2,202 retracted papers tagged as paper-mill products and an equal number of papers treated as presumed-genuine controls. They tested the model against independent material collected by image-integrity experts, where it reached 91% accuracy.
What the screen found
The team ran the classifier across 2,647,471 original cancer research papers indexed in PubMed and published from 1999 through 2024. It flagged 261,245 papers, or 9.87%, as having textual characteristics similar to retracted paper-mill work.
The proportion rose from roughly 1% in the early 2000s to more than 15% of annual cancer research output by the early 2020s. The signals also appeared in the top 10% of journals by impact factor and were especially common in fundamental research and studies of gastric, bone and liver cancer.
More than 170,000 flagged papers had affiliations with Chinese institutions, representing 36% of China-affiliated cancer papers in the dataset.
Why the numbers need caution
The model read only titles and abstracts. Its “genuine” controls were proxies rather than papers proven free of misconduct, and the authors acknowledged that some underlying labels could be wrong. A flag therefore indicates textual similarity, not proof that data were fabricated or that an author bought a paper.
The Scientist reported that independent digital-forensics researcher João Phillipe Cardenuto welcomed the screening approach but warned that the model could learn language patterns common among Chinese authors rather than features specific to fraud. The researchers included China- and Taiwan-affiliated controls to reduce that risk, but the country-level estimates still require caution.
The tool is best understood as a way for publishers and research-integrity teams to prioritize papers for closer review, not as a fraud detector that delivers a final verdict.
