Lancet audit: 4,046 citations to nonexistent works found across 2.5M biomedical papers

jonc101x · x · 2026-11-02

A 15-part X thread by Jonathan Klahr promotes a talk, "When AI Writes the Science: Fabricated References, Agentic Literature Synthesis, and Ambient AI," with Max Topaz of Columbia University, lead author of the Lancet audit of fabricated citations.
Key findings from the audit: it checked 97.1 million references across 2.5 million biomedical papers and found 4,046 citations to work that does not exist, spread across 2,810 papers. No publisher action was taken on 98.4% of them — fabricated references mostly stay in the published record uncorrected.
The talk covers three lines of work: fabricated references entering the literature; agentic literature synthesis (what AI agents can and can't do when summarizing an entire research field); and ambient AI in clinical documentation, including what these tools get wrong and practical guardrails for researchers, journals, and health systems.
The thread also flags the incentive problem: rigorous, validated information is far slower to produce than AI-generated slop, and it questions how much beyond citations in manuscripts and results is fabricated but hard to verify.
Meta-irony: the thread itself was prepared by Claude Cowork.

More from Safety