Are AI-fabricated citations a problem in medicine?
Yes — and the stakes are higher here than almost anywhere else. In medical and biomedical writing a single reference can underpin a dosage, a contraindication, or a clinical recommendation, so a fabricated source isn't just sloppy scholarship — it can mislead care.
What the evidence shows
Early experiments asking language models for medical references found high rates of fabricated and inaccurate citations, and a 2026 audit of millions of papers — including the peer-reviewed biomedical literature indexed in PubMed Central — documented a sharp rise in non-existent references after AI tools became widespread. Crucially, the audit found fabricated citations across the full range of journals, not only low-impact ones.
Why medicine is especially exposed
- Claims carry weight. A reference attached to a clinical statement implies evidence; a fake one manufactures evidence that was never there.
- Plausible specifics. Models readily generate real-sounding journals (NEJM, The Lancet), real author names, and DOIs in the right format — so a fabricated medical citation often looks impeccable.
- Retraction matters more. Citing a retracted trial as current support is its own hazard, distinct from fabrication — see retractions vs. fake citations.
Protecting a medical manuscript
Verify every reference against authoritative indexes (PubMed, Crossref) before submission: confirm the work exists, that the DOI resolves to it, that the authors match, and that it has not been retracted. For a reference list of any size, automate it — the check is mechanical and attention fades exactly where fabrications hide. See how to check a citation and how verification works.