A new study finds an alarming increase in the number of fabricated citations in published scientific papers. Are AI hallucinations to blame?gettyThe number of fabricated citations included in published research studies is a rapidly growing problem, finds a new peer-reviewed research letter published last week in The Lancet. Investigators are relying more often on AI to help write and edit the papers they submit to journals for publication. That practice and the risks of large language models "hallucinating" a title that does not correspond to a real citation is the likely culprit behind the increase, according to the Columbia University team that conducted the study. Lead author Maxim Topaz, the Elizabeth Standish Gill Associate Professor of Nursing at Columbia University, said he decided to conduct the study after his own use of AI to help edit a paper he was writing resulted in the addition of a fabricated citation that was later flagged by editorial staff. Embarrassed by that experience, Topaz and his colleagues investigated how widespread the problem of bogus citations might be. They used an automated citation verification system to audit more 2.5 million papers that were published between January 2023 and February 2026 and included in PubMed Central’s Open Access dataset.Those papers contained more that 125 million bibliographic references, of which 97.1 million (77%) carried a PMID and were verified; the remaining 23% mainly referred to websites, books, and other studies and were excluded from the analysis.MORE FOR YOUSeveral filters were applied to identify false positives that might arise from title artifacts like abbreviations or misspellings. The references that passed those initial filters were then verified against PubMed, Crossref, OpenAlex, and Google Scholar. If a citation could not be found in any of those four databases, it was classified as a fabricated reference.The team found a total of slightly more than 4,000 fabricated citations among 2,800 papers. That’s not that many, but the bad news is that hallucinated citations are clearly on the rise, increasing roughly twelve-fold in just three years. In 2023, approximately one in 2,828 papers contained at least one fabricated reference. By 2025, that number had risen to one in 458, and in the first seven weeks of 2026, one in 277 papers included at least one fabricated reference. Typically, the fabricated references “were not obviously defective,” according to the authors, making them hard to detect. They dealt with specific scientific topics, were correctly formatted, attributed to real researchers, and carried plausible publication dates. The authors noted that their sharp increase after 2024 coincided with the expected publication lag following the time when LLM began to be widely used in late 2022 and 2023. The implications of the findings are serious, and they are not an isolated discovery. Last month Nature reported that tens of thousands of publications from 2025 “might include invalid references generated by AI.” Fabricated references can result in false leads, and they can lend a spurious boost to a study’s credibility. Referencing papers that don’t exist suggests a shoddy approach to research and may undermine the confidence the public has in the integrity of the scientific process. Their inclusion in a published study also reflects poorly on the peer review process that resulted in the study’s acceptance for publication. What Can Be Done?The authors recommended that journal publishers use automated reference verification in submission workflows before peer review occurs and that indexing services add integrity metadata to article records so that subsequent users can assess reference reliability. In addition, they believe that publishers should retroactively screen existing publications and issue corrections or retractions when fabricated references are discovered that compromise a paper’s conclusions. Finally, they suggest that establishing a category for fabricated references in major research integrity databases would improve tracking and accountability. In a press release, Topaz said that the findings of the new study could ultimately have an impact on patients because medical professionals make treatment decisions based on research-influenced clinical guidelines, adding that “a medical professional or clinical guideline developer has no way of knowing that the evidence they are relying on does not exist. For example, one paper we reviewed had 18 out of 30 fake references. Some of those citations are already being cited by other papers and appear in systematic reviews that inform clinical care.”In an accompanying comment published in The Lancet, Howard Bauchner (professor of pediatrics and public health at Boston University) and Frederick Rivara (professor of pediatrics at the University of Washington School of Medicine) maintained that it was ultimately up to authors themselves to take responsibility for the accuracy of their reports, including the references. “Given that public trust in science appears to be waning in countries around the world, renewed efforts are needed to enhance research integrity,” they wrote.