Audit Finds Fabricated Citations Across Biomedical Literature

The correspondence by Maxim Topaz and colleagues, published in The Lancet, reports an AI-assisted audit of approximately 2.5 million biomedical papers drawn from the PubMed Central Open Access subset. The authors say they verified 97.1 million structured references and identified 4,046 references the team characterized as fabricated, appearing across 2,810 papers, per The Lancet correspondence. The correspondence and reporting in outlets including Nature, CIDRAP, CBS News, and Retraction Watch describe a roughly 12-fold increase in fabrication rates from 2023 to 2025, with an early-2026 rate of about one in 277 papers, the audit authors report.

The audit team used an automated reference-verification pipeline that compared citing metadata with bibliographic records retrieved from PubMed, CrossRef, OpenAlex, and Google Scholar, and applied filters to reduce false positives. CIDRAP reports the system achieved about 91% precision after filtering. The team labeled a reference as fabricated when no corresponding record was found in those reference databases. Nature corrected initial coverage to note the study screened the PubMed Central Open Access subset rather than the entire PubMed index.

Industry observers: Automated verification across tens of millions of references can scale detection but depends on the scope and coverage of the underlying databases. The audit screened the PMC Open Access corpus, which omits paywalled content; therefore the measured rates describe that subset and may not directly generalize to the full published literature. Precision reported by the authors implies some residual false positives remain; reported filtering and human checks mitigate but do not eliminate classification uncertainty.

Reporting across Retraction Watch, CIDRAP, and Nature cites the correspondence warning that fabricated references can originate from multiple sources, including paper mills, intentional misconduct, and uncritical use of large language models. Retraction Watch and CIDRAP note the sharpest increase in fabricated references beginning in mid-2024, a period when public use of LLMs and AI writing tools grew. CBS News quotes Maxim Topaz saying, "Your doctor could be making decisions around treatment based on studies that never existed," highlighting concerns about downstream impacts on clinical guidelines and patient care.

Industry observers: The finding joins a wider conversation about metadata integrity and provenance in scholarly publishing that predated generative-AI adoption but now intersects with LLM failure modes such as hallucinated citations. Publishers and journals quoted by Retraction Watch and other outlets report investing in integrity screening tools and publication-ethics workflows; PLOS told Retraction Watch it is "exploring options for system-wide reference integrity screening," per Retraction Watch.

• •Whether journals and publishers expand automated reference-checking to cover submission pipelines and postpublication audits.
• •Development and adoption of standardized APIs or services that can verify DOIs and metadata at scale across paywalled and open content.
• •Follow-up studies that replicate the audit across broader corpora or apply different verification heuristics to quantify false-positive and false-negative rates.

Industry observers: Data scientists, journal editors, and research-software engineers should treat reference verification as a measurable metadata quality problem amenable to tooling. Integrations that cross-check claimed identifiers against multiple bibliographic sources and flag mismatches can reduce propagation of fictitious citations, but they require dataset coverage and calibration to minimize nuisance alerts.

The Lancet correspondence and subsequent reporting indicate that, at the time of the audit, most papers with flagged references had not been corrected or retracted, per Retraction Watch and the audit authors' summary in The Lancet.