ai
Study Detects Fabricated References in 0.11% of Papers, Rising with AI Adoption
Researchers identified a sharp rise in fabricated citations in scientific papers from 2023 to early 2026, attributing most to AI hallucinations.
May 30, 6:19 PM·1m read1 source
ForbesAudio version
Tap play to generate a narrated version.
Developing·Limited corroboration so far. This page will refresh as more sources emerge.
A research team identified 4,046 fabricated references across 2,810 published scientific journal articles between January 1, 2023, and February 18, 2026. The team reviewed 2,471,758 papers containing 125,615,773 references in PubMed Central’s Open Access collection.
They developed an automated reference verification system that compared citations listed in papers with actual bibliographic records in PubMed, Crossref, OpenAlex, and Google Scholar.
5 Haiku from Anthropic to classify discrepancies as honest errors or fabrications. Any reference that could not be found in the databases was considered fabricated. In 2023, approximately one in 2,828 papers had at least one fabricated reference.
In 2025, the rate rose to one in 458. In the first seven weeks of 2026, the rate reached one in 277. The rate of papers with fabricated references increased over 12-fold from 2023 to early 2026. One paper published in 2025 in an open access oncology journal covering ureteroileal anastomotic techniques had 18 fabricated references out of 30 total citations.
Sixty percent of the citations in that paper were presumably AI-fabricated. Two hundred forty-six different papers contained three or more fabricated references. A pair of authors published 11 different papers in a single surgical journal in 2025 that together contained 15 fabricated references.
Review articles had fabricated reference rates 57 percent higher than other types of papers. 6 per 10,000 for non-review papers. The research team attributed most fabricated citations to hallucinations produced by large language models.
The increase in fabricated citations corresponded with greater use of generative AI platforms such as ChatGPT, Claude, and Perplexity. The research team members are Maxim Topaz, Nir Roguin, Pallavi Gupta, and Zhihong Zhang of Columbia University, and Laura-Maria Peltonen of the University of Eastern Finland. They published their findings as a correspondence in The Lancet.
