TL;DR
- Audit Rate: A recent audit found one in 277 biomedical papers carried one or more non-existent references in early 2026.
- Growth Trend: The reported fake-reference rate rose from one in 2,828 papers in 2023 to one in 458 in 2025.
- Clinical Stakes: Because physicians already use AI for research summaries and documentation, citation checks now matter well beyond journals.
An audit across 2.5 million biomedical papers published by The Lancet has found that one in 277 biomedical papers carried one or more non-existent references.
Maxim Topaz, associate professor at Columbia University’s School of Nursing, said in one case an AI-assisted draft had inserted a fabricated source into one of his papers before publication.
Because the audit covered nearly 2.5 million papers and 97 million citations, the finding extends beyond one researcher’s mistake.
Researchers found more than 4,000 fabricated references across nearly 3,000 biomedical papers, and Topaz said 98.4% had not been retracted when he checked them. That leaves journals, reviewers, and clinicians to clean up bad citations after publication instead of stopping them during review.
False references matter in medicine because one bad citation can move from a manuscript into later reviews, summaries, and treatment guidance.
“This is the evidence chain, that’s how we care for and treat people. If you put the fictional study at the bottom of the stack, the whole structure inherits it.”
Jonathan Topaz, associate professor at Columbia University’s School of Nursing (via Fortune)
How One Citation Error Became a Measured Warning
The available evidence does not justify saying every false reference came from AI, but it does show how quickly verification burdens grow once an invented citation reaches the literature.
In 2023, just one in 2,828 papers contained one or more fake references. By 2025, one in 458 papers did, and the reported rate had grown more than 12-fold across the period.
That sequence points to a widening verification gap before the 2026 spike. Topaz warns that delayed checks make cleanup harder once false citations spread through the literature.
The Growth Curve and the Verification Gap
A fabricated citation to one of Guillaume Cabanac’s papers in a 2026 dentistry paper showed how easily a made-up source can pass into published literature. Nature warned that tens of thousands of 2025 publications might include invalid references generated by AI, so the current audit may describe only part of the cleanup burden.
Another high-stakes version of the hallucination risk surfaced in the 2025 FDA ‘Elsa’ controversy over hallucinated studies. Publishers are also responding to unchecked AI output through the ArXiv enforcement response and the legal-citation failure in Anthropic’s copyright case.
Editorial controls still offer a direct intervention point because manual verification can catch fabricated references before publication. Editors get their clearest chance at the manuscript stage, when a missing source can still be isolated before it enters databases, literature reviews, and later summaries.
What Journals Can Still Catch Before Publication
Editors can still stop the damage chain because manuscripts can be rejected or post-acceptance approvals rescinded when fabricated references are identified. Copy editors, peer reviewers, database checks, and author-side verification all matter because the problem starts in drafting but becomes harder to unwind after publication and indexing.
Because eighty-one percent of physicians used AI in their practices in 2026, verification pressure now reaches beyond academic publishing. Common physician uses already include medical research summarization and clinical documentation, extending citation-check pressure into everyday care workflows.
That wider adoption does not prove doctors are publishing fake references in routine care. It does show why evidence quality matters more once AI-generated summaries sit closer to treatment, triage, and recordkeeping.
“The key lesson is not to avoid AI altogether, but to ensure careful human verification of all outputs.”
Lisa A. Fortier, Editor-in-Chief of JAVMA and AJVR (via American Journal of Veterinary Research)
Journal editors still control the last practical checkpoint before an invented source becomes part of the published evidence chain. Topaz said in The Scientist that even experienced AI users still need safeguards because “it can happen to anyone”.
