A new Lancet study led by Maxim Topaz at Columbia University’s School of Nursing found fabricated references buried in biomedical papers, highlighting how AI hallucinations can enter the scholarly record. Researchers audited nearly 2.5 million biomedical papers and 97 million PubMed Central citations and identified more than 4,000 fake references across nearly 3,000 papers. Topaz said the risk appears to have accelerated after 2024, when AI tools became more widely used in research workflows. The team reported that the number of papers containing at least one fake reference rose from one in 2,828 in 2023 to one in 458 by last year, and that in the first seven weeks of 2026, one in 277 papers included non-existent references. The study underscores a compounding failure mode in medicine: clinical trials cite earlier studies, systematic reviews aggregate trials, and guidelines cite reviews. A fabricated source early in that chain can therefore propagate downstream to clinicians and patients. The findings also drew attention to the practical editorial reality that grammar- and formatting-support tools may silently introduce incorrect citations, suggesting publishers and institutions may need additional reference verification controls for AI-assisted manuscript preparation.