In a preprint examining 270 medical reviews written by language models, 76,8% of incorrect references pointed to a real but different article

In a preprint examining 270 medical reviews written by language models, 76,8% of incorrect references pointed to a real but different article

On September 4, medRxiv, a server for medical preprints, published a study of 270 cardiology reviews. Three language models with web search enabled wrote the reviews, and the authors checked 8 050 references against PubMed, a database of medical publication records.

Each article indexed in PubMed receives a PMID, the number identifying its record. In one review, a model attached PMID 17565084 to a reference for an article about cardiac resynchronization therapy. However, that PubMed record leads to a JAMA article about National Institutes of Health grants for early-career physician-scientists. Checking a reference requires first matching the identifier to the named article, then determining whether the article supports the statement for which it was cited.

The three models each wrote ten reviews for each of nine cardiology topics; each review contained 600–900 words and 30 references. Among the 8 036 references with a PMID or DOI (a persistent digital identifier for an article), 11,5% were problematic for GPT-5.5, 29,2% for Claude Opus 4.8, and 29,6% for Gemini 3.5 Flash. Most problems involved mismatched references: a valid PMID or DOI was attached to a different article. These accounted for 76,8% of errors.

“A reference combines several claims about an article: its title, authors, journal, year, PMID, DOI, and the statement it is meant to support.”

The authors also applied CoVe, a chain-of-verification procedure in which a language model breaks a reference down into individual questions. It retrieves the PubMed record using the supplied identifier, separately searches for the article by title and first author, then compares the authors, journal, year, volume, and pages. In a sample of 270 references labeled by a cardiologist, CoVe detected 60 of 62 problems, including all 51 mismatched references and all four fabricated articles; the procedure incorrectly flagged three valid references as problematic.

The authors separately compared nine statements from the reviews with the abstracts of the cited articles. Checking a reference involves two steps: the identifier and title must point to the same article, and that article must support the statement in the text.