Audit of 2.5 Million Biomedical Papers Finds 4,046 Fabricated Citations
A Lancet correspondence published May 9, 2026 reports that researchers identified 4,046 fabricated references embedded in 2,810 biomedical papers. The audit covered 2.5 million open-access papers in PubMed Central spanning January 2023 through February 2026, examining 97.1 million verified references.
The fabrication rate has accelerated sharply. In 2023, roughly 4 fabricated references appeared per 10,000 papers. By early 2026, that rate reached 57 per 10,000 papers - a more than 12-fold increase in less than three years.
How the audit worked
Researchers built an automated pipeline that compared cited references against bibliographic databases including PubMed, Crossref, OpenAlex, and Google Scholar. Mismatches were flagged for manual review using an AI assistant to triage cases at scale.
References that could not be located in major databases were classified as fabricated. The methodology has limits: it may miss legitimate but poorly indexed material, and the audit covered only open-access papers, not the full PubMed corpus.
AI writing tools linked to the rise
The correspondence attributes the sharp increase since mid-2024 to growing use of generative AI writing tools in research workflows. Prior studies estimate that 30 to 69 percent of references generated by large language models in biomedical contexts may be fabricated.
Fabricated citations can result from paper mills, deliberate misconduct, or researchers using AI writing assistants without critical review of the generated text.
Why this matters for practitioners
Fabricated citations corrupt the scientific record that researchers rely on for literature reviews and model training. They also propagate downstream through knowledge graphs and retrieval systems that ingest published papers, potentially contaminating training data for biomedical NLP models.
Teams assembling training corpora or systematic reviews should assume that a measurable portion of published biomedical references may be fabricated. Automated screening combined with spot checks may be necessary to reduce contamination.
Recommended actions
The authors recommend three steps:
- Publishers should verify references at manuscript submission.
- Indexing services should add metadata flags for suspected fabricated references.
- Research-integrity databases should create a dedicated category for fabricated citations.
What to monitor
Watch whether major publishers implement automated reference checks at submission. Track whether PubMed, Crossref, and OpenAlex add metadata fields to flag suspected fabrications.
Follow-up audits across broader corpora beyond open-access papers will clarify the scope of the problem. Emerging standards for documenting AI assistance in manuscript preparation may also help separate honest errors from fabrications.
Researchers working with published biomedical literature should treat reference verification as part of data preparation, not an optional step. The faster adoption of Generative AI and LLM tools in writing makes this more urgent.
Your membership also unlocks:
