A reference list is supposed to be the most boring part of a paper, and that is exactly why it matters. It tells readers where the writer’s ideas came from and lets other people check the evidence for themselves. But AI tools can now invent papers that were never written. In a Nature report published on May 14, 2026, researchers who audited 111 million references across arXiv, bioRxiv, SSRN, and PubMed Central estimated that papers and preprints from 2025 alone contained 146,932 hallucinated citations. The problem was strongest in fields with rapid AI use, and Nature said the highest level appeared on a social-sciences preprint site. (nature.com)
This is not only a preprint problem. A separate Nature report, covering a Lancet study published on May 7, 2026, said that an audit of 2.5 million biomedical papers found 4,046 fabricated references across 2,810 papers. The researchers checked whether a reference’s title matched the paper linked by its DOI or PubMed ID. That means some citations looked real at first glance, but the details did not actually lead to the claimed study. (nature.com)
Why is this serious? Because many readers trust a citation without opening it. If a fake source slips into one paper, later writers may copy it, and the error can spread. The large arXiv study warns that preprint moderation and journal publication catch only a fraction of these mistakes. It also suggests that hallucinated citations can give extra credit to scholars who are already well known, and more often to male scholars, which could make old inequalities even stronger. (arxiv.org)
Can trust still be protected? Yes, but not by AI alone. Springer Nature says clearly that authors, not AI, are responsible for the accuracy, originality, and integrity of a manuscript, and that AI use must be declared. The ICMJE, whose recommendations shape many medical journals, says humans must review AI-generated content carefully, provide proper citations, and disclose AI assistance. Crossref also offers free tools to match references with DOIs. In short, smart software can help, but the final guard of research trust is still a careful human being. (group.springernature.com)










