Fake references in medical papers skyrocket, linked to AI tools

A large-scale check of 2.5 million biomedical papers found fake references are becoming a serious problem. Since 2023, fake references have increased more than twelve times. Researchers from Columbia University and other groups published a report in The Lancet. They looked at 2.47 million papers from January 2023 to February 2026, checking 97.1 million references. They found 4,046 fake ones in 2,810 papers. A reference was considered fake if its title couldn't be found in PubMed, Crossref, OpenAlex, or Google Scholar.

The trend is clear: in 2023, about 4 fake references appeared per 10,000 papers. By late 2025, that rose to 51.3 per 10,000 papers, and in early 2026, it hit 56.9 per 10,000 papers. The team thinks this is linked to ChatGPT and similar AI tools that became popular in late 2022. These fake references are tricky—they fit the paper's topic, look properly formatted, and use real author names. Review papers are hit the hardest, with 57% more fake references than other papers. Since review papers often guide medical decisions, fake references could mess up treatment advice.

Efforts to fix this are uneven. arXiv now bans authors for a year if they use AI to make up references. But even top AI conferences like NeurIPS can't easily spot fake citations. The open-source tool CiteAudit helps, but tests show commercial AI models can't catch their own fake references. The team suggests checking references automatically before peer review, adding honesty metadata to paper databases, scanning published papers again, and creating a special category for fake references in research integrity databases.