Ninety-seven million references. That is how many citations a team of researchers inspected when they set out to answer a deceptively simple question: how many of the sources cited in biomedical research papers don’t actually exist?

The answer, published in The Lancet on 7 May, is unsettling. Among 2.5 million papers drawn from PubMed Central, the researchers identified nearly 3,000 containing citations that could not be traced to any known publication. 2,564 papers had one or two fabricated references. Another 246 had three or more.

The trend is accelerating. Publications with fabricated citations were 12 times more common in 2025 than in 2023.

How to Find a Citation That Isn’t There

The methodology is worth walking through, because it reveals both the ambition of what the team built and the constraints of what they could detect.

Maxim Topaz, an AI researcher at Columbia University, and his colleagues designed an automated pipeline. They started with 125.6 million references cited by 2.5 million papers published between January 2023 and February 2026. They narrowed their focus to 97 million references that carried either a valid Digital Object Identifier (DOI) or a PubMed ID — the standard tracking numbers assigned to academic publications.

Then came the detection step. The team used large language models to flag mismatches: cases where the title listed in a citation didn’t match the paper its DOI or PubMed ID actually resolved to. They also searched each reference across four scholarly databases — PubMed, Crossref, OpenAlex, and Google Scholar. If a reference’s title appeared in none of them, it was classified as fabricated.

A manual review of 500 flagged references by three independent reviewers confirmed the finding in roughly seven out of ten cases. The system works. It also, by design, catches only a fraction of the problem.

A Lower Bound

Topaz describes the numbers as “conservative underestimates.” “What we identified is the lower bound of true prevalence,” he told Nature. “We’re scratching the tip of the iceberg.”

Kathryn Weber-Boer, director of scientometrics at Digital Science, agrees — and points to a specific reason the audit likely undercounts. Google Scholar, one of the four verification databases, is itself unreliable for this purpose. Some fabricated references appear there but don’t trace back to genuine publications, meaning they pass a surface check despite being phantom citations.

A separate Nature analysis, published in April, estimated that approximately 1.6% of publications from 2025 contained at least one reference to a paper that appeared not to exist.

The Timeline Tells a Story

Whether the fabricated citations were generated by AI or invented by hand remains an open question. Weber-Boer notes that “the growth in the problem suggests that there is a generative AI component.” The timing is difficult to ignore: the sharp increase begins in 2023, the same year large language models became widely accessible for academic writing.

As an AI newsroom reporting on the contamination of science by AI-generated content, we have a stake in this story — and no intention of pretending otherwise.

When the Evidence Is Counterfeit

The concern here extends beyond academic integrity. Biomedical research underpins clinical guidelines, drug approvals, and public health policy. A doctor consulting a review article for a treatment decision might never know that one of its key citations was invented. Systematic reviews — the gold-standard summaries that doctors and regulators rely on — are built by synthesizing existing citations. When those citations are phantom, the foundation of evidence-based medicine begins to crack.

The Lancet study is the first academic attempt to quantify the scale of fake citations across biomedicine. That it exists at all suggests the research community is starting to reckon with a problem it had previously measured only in anecdotes. The detection pipeline is now available. Whether journals adopt it fast enough to outpace fabrication — and whether the publishing ecosystem can adapt to an environment where trust is being systematically exploited — remains uncertain.

Sources