Columbia University audit finds AI citations eroding biomedical trust

An audit of 2.5 million biomedical papers by researchers at Columbia University and other institutions points to a blunt failure mode in scientific publishing: fabricated references are no longer a minor edge case. According to The Decoder, the rate of fake citations has increased more than twelvefold since 2023.

That matters because biomedical research is not built only on individual findings. It is built on a chain of verifiable claims. A reference is one of the basic safeguards in that chain: a reader, reviewer, editor, or guideline author must be able to open the source and check what was actually shown. If the cited paper does not exist, and still looks plausible enough to pass through the system, the problem is not cosmetic. It is a fault in the trust infrastructure.

The researchers suspect a connection to the widespread use of language models. The reason is not only the timing after 2023, but the shape of the errors. The fake references reportedly match the topic of the paper, follow normal citation formatting, and look like something that could plausibly exist. In other words, they are not random nonsense that immediately breaks the page. That is exactly what makes them dangerous: the system can treat them as ordinary noise until someone tries to verify the underlying source.

The most sensitive part is the connection to papers that shape clinical guidelines. Clinical recommendations are not usually changed by a single citation, but they are built from literature reviews, meta-analyses, and networks of prior work. If non-existent sources enter that network, the risk is not that every affected paper instantly alters care. The risk is slower and more corrosive: polluted literature that makes verification harder, wastes editorial and reviewer time, and weakens confidence in papers that need to be unusually precise.

The most uncomfortable figure in the reported findings is that 98 percent of the affected papers have received no response from their publishers. That suggests the correction process is not moving at the same speed as the problem. For publishers, databases, and editorial teams, this can no longer be treated as a matter of author discipline alone. Reference checking needs to become a machine-assisted editorial routine, comparable to DOI, indexing, and metadata checks through services such as Crossref or biomedical databases such as PubMed.

AI is not a magical explanation for every bad citation, but it is an accelerator of an old failure. Authors could always misquote, copy references without checking them, or move work through weak editorial pipelines. The difference is that language models can generate a convincing, topic-aligned bibliographic artifact in seconds. If nobody verifies it, the invented citation begins to look like part of the academic landscape.

For biomedicine, the tolerance threshold should be low. A paper that cannot prove its own sources should not quietly travel toward review articles, guidelines, and clinical practice. The next test for publishers is not a broad statement against hallucinated references, but a measurable response: how quickly they can find, flag, and correct papers where the bibliography has stopped being evidence and has become fiction.