When AI-fabricated citations reach the record — the documented cases
A running record of cases where fabricated references — invented papers with plausible authors, journals, and DOIs — reached published research, survived peer review, or entered a court filing. We didn't find any of these. Each one was documented by someone else: a journal, a research group, an integrity tracker, or a judge. This file collects the public record and points you at the primary source for every entry.
Every case below was found, measured, reported, or ruled on by a third party. Cento compiles cases that others have already made public and links each to its primary source; we make no original findings of misconduct and name no person we have not seen named in a published source. Where an entry rests on a preprint or a single analysis, we say so. The point isn't to accuse anyone — it's to show, from the public record, that fabricated citations are now a documented problem in the literature.
In the published literature
Fabricated references that made it past submission and into indexed papers.
Fabricated references in biomedical papers rose about 12-fold in three years
An audit of nearly 2.5 million PubMed-indexed biomedical papers checked 97.1 million references and identified 4,046 fabricated citations across 2,810 papers. The share of papers carrying at least one non-existent reference climbed from 1 in 2,828 in 2023 to roughly 1 in 277 in early 2026, with the sharpest rise in mid-2024 as generative-AI writing tools spread. In one flagged paper, 18 of its 30 references appeared to be fabricated.
Nearly 147,000 non-existent citations across four repositories in a single year
An audit of 111 million references across 2.5 million papers on arXiv, bioRxiv, SSRN, and PubMed Central produced a conservative estimate of 146,932 hallucinated citations in 2025 alone. The fabrications clustered in fields with rapid AI adoption and in papers carrying the linguistic signatures of AI-assisted writing — evidence that the pattern tracks tool use, not chance.
In peer review
Fabricated citations that expert reviewers didn't catch.
About 100 fabricated citations were accepted at NeurIPS 2025
A preprint analysis identified roughly 100 AI-generated, non-existent citations across 53 accepted papers — about 1% of the conference — at one of the most competitive venues in machine learning. Each paper had been read by three to five expert reviewers, and the fabrications survived review anyway. Peer review checks methods and novelty; it rarely checks whether every cited paper exists.
In court
The same failure, outside science — where a ruling put it on the public record.
A U.S. court sanctioned lawyers for a brief built on ChatGPT-invented cases
In Mata v. Avianca, two attorneys submitted a filing citing judicial opinions that did not exist; the citations had been produced by ChatGPT, which then assured them the cases were real. Judge P. Kevin Castel of the Southern District of New York imposed a $5,000 sanction. It remains the clearest public illustration that a fabricated citation can look flawless and still be entirely invented.
A public database now tracks more than 1,600 court rulings that flag AI-hallucinated citations
Mata v. Avianca was not a one-off. A continually updated database catalogues legal decisions in which courts found that a party relied on AI-hallucinated content — most often fabricated case citations. It had passed 1,600 documented rulings across multiple countries by mid-2026, a scale that marks the failure as systemic rather than anecdotal.
Measured in controlled tests
Studies that asked AI tools for references and checked how many were real.
In ophthalmology, roughly a third of AI-supplied references couldn't be verified
A study in JAMA Ophthalmology asked two versions of a widely used chatbot to generate ophthalmic scientific abstracts with references. The mean rate of non-verifiable — fabricated — references was 33% for the earlier version and 29% for the updated one. The abstracts read well; the references frequently pointed to papers that did not exist. This is Cento's home specialty, and the reason we start here.
47% of references in AI-generated medical content were fabricated
Researchers examined 115 references produced by ChatGPT for medical content. Only 7% were both authentic and accurate. 46% were authentic but cited inaccurately, and 47% were fabricated outright — papers that were never written. Put another way, more than nine in ten references had a problem.
Hallucination rates ran from 29% to over 90%, depending on the model
A comparison testing chatbots on systematic-review references reported a 28.6% hallucination rate for GPT-4 and 91.4% for Bard. Newer models fabricate less than older ones — but "less" is not "never," and the gap between them shows fabrication is a property of the approach, not a bug one version fixed.
What the cases have in common
Read together, these cases point to one mechanism. A fabricated citation is convincing because a language model is built to produce plausible-looking text: a real-sounding author string, a title-shaped phrase, a journal that co-occurs with the topic, a well-formed DOI. The format is perfect because format is the only thing being modelled — whether the paper exists is not a question the model can ask. The failure shows up in a court filing and in a clinical abstract for the same reason.
Every case here was caught after the fact — by a judge, an auditor, a reviewer reading closely. The only way to remove the risk is to never let the model choose a reference in the first place.
That is the design choice behind Cento: retrieve real papers first, constrain the model to cite only from that retrieved set, and validate every citation before it reaches you. Fabrication isn't caught later — it can't occur. See AI citation fabrication: the data and the fix for the full mechanism, or how to check whether a citation is real if you want to vet references yourself today.
Frequently asked
Has AI actually caused fake citations in published research?
Yes, and it has been measured. A 2026 Lancet audit of nearly 2.5 million biomedical papers found the share carrying a non-existent reference rose to roughly 1 in 277 in early 2026, about twelve times the 2023 rate. A preprint analysis found about 100 fabricated citations across 53 papers accepted at NeurIPS 2025, and U.S. courts have sanctioned lawyers over AI-invented case law. Each was documented by a third party.
Who documents these cases?
Journals and their research-integrity teams, independent research groups, retraction trackers such as Retraction Watch, science reporters, and courts through published rulings. This file aggregates cases those parties have already made public and links each to its primary source. Cento makes no original findings.
How is this file kept current?
We review the public record weekly for newly documented cases. A case is added only when it has a primary source we can link, the source resolves, and the finding is attributed to whoever made it. Cases we can't yet verify to a primary source are held back, not published.
Write papers that can't join this list
Cento is the AI co-writer that can't fabricate a citation. Built for medical researchers, starting in ophthalmology.
Join the waitlist