The Citation Crisis Is Concentrated in Two Publishers — and the Problem Is Getting Worse
The contamination is not random. A new Lancet study auditing 2.5 million biomedical papers and 125.6 million references found 4,046 fabricated citations — and more than a third of them trace back to just two publishers. The highest-rate publisher generated fabrications at 14 times the rate of the most selective journals in the dataset. The rate itself is also climbing: from roughly 4 per 10,000 papers in 2023 to 56.9 per 10,000 in early 2026 — a more than twelvefold increase with no historical precedent.
The study, led by Maxim Topaz, a nurse and health AI researcher at Columbia University, verified every flagged reference against four independent databases: PubMed, Crossref, OpenAlex, and Google Scholar. During the first seven weeks of 2026 alone, the rate reached 1 in 277 papers. The problem has also caught the attention of Nature, which reported in April 2026 that tens of thousands of publications from 2025 might include invalid AI-generated references.
"This is one of the first papers that's telling us something about the quality of what's being produced with LLMs, and it's a signal of slop," said Misha Teplitskiy, a sociologist of science at the University of Michigan who has studied AI use among academics but was not involved in the paper.
Topaz came to the problem the way many researchers now do: through personal experience. He used an AI chatbot to help edit an editorial, checked the citations, and still nearly published a reference to a paper that did not exist. "I was deeply embarrassed: I checked for that, and it still almost happened to me. This is how I ended up thinking about other people."
The fabricated citations are engineered to look legitimate — real-sounding authors, plausible journals, credible publication years. A side-by-side comparison on the CITADEL dashboard that Topaz published alongside the paper shows a fabricated citation and a real one appearing equally professional. Only automated verification across multiple databases can reliably distinguish them. The worst case found so far: a single published paper on ureteroileal anastomotic stricture rates in which 18 of 30 references pointed to studies that do not exist.
"The more serious issue here is that citation practices are changing with the generative AI use," said Mohammad Hosseini, a professor at Northwestern University who studies research integrity. "Previously you would read a paper and then make notes on it. Once you were writing a paper, you would be like, 'Is that paper or book relevant?' It was a much more reflective process, whereas now people simply use their hunches to prompt ChatGPT or other AI tools, and then they have a bunch of citations that they can sprinkle over their papers. That means that the engagement with the literature is becoming increasingly more superficial, and that is neither good for the researcher, nor for society, nor for our publication practices."
"It shows how fast they want to get published and how desperate they are to get published, which is a reflection of a flawed scholarly evaluation model that puts so much emphasis on peer reviewed publications," Hosseini said.
Topaz puts it more bluntly: the presence of fabricated citations signals that authors who use them did not bother to spend half an hour checking their references.
The contamination has consequences beyond the papers themselves. Systematic reviews and clinical guidelines that attempt to build on published literature can inherit these fake references, propagating errors into the evidence base that informs medical practice. A widely publicized example from 2025: the White House MAHA report on chronic disease contained several incorrect citations that onlookers suspected were generated by artificial intelligence. More recently, South Africa withdrew its debut national AI policy less than three weeks after publication after investigators found parts of the document contained fabricated academic references. Fabricated citations have also appeared in U.S. government reports requiring corrections and in professional consulting outputs, according to an arXiv preprint on compound deception in elite peer review. That study found that at least 53 accepted papers at the 2025 NeurIPS conference — roughly 1% of the 5,290 submissions reviewed — contained fabricated citations that evaded three to five expert peer reviewers each.
Top-tier journals say they are not seeing the problem. Representatives for the New England Journal of Medicine, JAMA, and the Science family of journals told STAT they use automated reference validation tools and hold authors accountable for the accuracy of their citations. Science noted it has yet to encounter a published paper with a fabricated reference. PLOS, which operates in the open-access, author-pays model most associated with the problem, said it has seen "numerous" unverifiable references in submissions and is exploring new workflows to address the issue, though it noted high false-positive rates in available tools.
The CITADEL dashboard is itself a partial answer to the problem — a free, automated citation audit tool that any journal or editor could run. But the incentive to use it, or to invest in deeper reference checking, is weakest at the publishers where the problem is most concentrated.
That is the uncomfortable underlying story here. AI did not create the pressure to publish fast and cheap. It amplified it. The technology makes it easier to generate plausible-sounding citations in seconds. What it cannot do is verify them — and the parts of the scholarly ecosystem with the least capacity or will to verify are precisely the ones growing fastest.