Overview
- The AI-assisted audit, published Thursday in The Lancet, scanned nearly 2.5 million PubMed Central papers and verified 97.1 million references, flagging 4,046 fabricated citations across 2,810 articles.
- Prevalence climbed more than twelvefold from 2023 to early 2026, rising from about one in 2,828 papers to one in 277, with the sharpest jump starting mid‑2024 as large language models gained wider use in writing and editing.
- The team matched reference titles to their DOIs or PubMed IDs and then searched PubMed, Crossref, OpenAlex, and Google Scholar, with a 500‑record check confirming fabrication in seven of ten flagged cases and noted limits that likely make the estimate conservative.
- At the time of the February audit, over 98% of affected papers showed no publisher action, while outlets like Taylor & Francis cited new checks and PLOS said it is exploring system‑wide screening and does not equate a single fabricated reference with misconduct by default.
- Researchers warned that fake citations are already appearing in reviews and could shape clinical guidance, and they urged automated pre‑review checks, integrity metadata in indexes, retroactive screening, and tracking of fabrications, noting clusters suggestive of paper‑mill activity and higher rates in some large open‑access journals.