A fabricated citation is a reference to a paper that does not exist. AI tools produce them because a large language model writes by predicting plausible text from patterns it learned, and a reference is one more kind of text it can generate without ever looking anything up.
Fabricated citations can look flawless. They have realistic author names, a fitting title, a real journal and a believable year, which is exactly why they slip past readers.
What counts as a fabricated citation?
A citation is fabricated when the work it describes was never published. That covers a few forms:
- Fully invented: the authors, title and journal details describe no real paper.
- Composite: real pieces are stitched together, such as real authors with a title they never wrote, or a real journal with a volume and pages that hold a different article.
- Borrowed identifier: a real PubMed ID (PMID) or digital object identifier (DOI) is attached to an invented title, so the identifier works but leads to a different paper.
In each case, a reader who tries to find the cited paper cannot, because it does not exist in the form described.
How is a fabricated citation different from a misattributed one?
A misattributed citation points to a real paper that does not support the claim it is attached to. The paper exists and the details may be perfect, but the study looked at a different population, measured a different outcome, or reached a weaker conclusion than the sentence suggests.
The two failures call for different checks. A lookup catches a fabricated citation, since the search comes back empty or finds a different paper. A misattributed citation passes the lookup, so you only catch it by reading the source against the claim. Both leave the reader believing a claim has evidence behind it when it does not. The subtler case gets its own post: why a real citation can still be the wrong citation.
Why do AI tools make up references?
A large language model is trained on a very large body of text. From that text it learns patterns: which words tend to follow which, how arguments are built, what a reference list looks like. When it answers, it builds the response piece by piece, each time choosing text that fits what came before.
References are among the most regular patterns in scientific writing. They follow fixed formats, reuse the same journals, and have titles that describe their findings. Ask a model for a source on a topic and it can produce text in exactly that shape: plausible surnames, a title that mirrors your question, a journal that publishes in the area, a year and a page range. The output looks like a citation because it was built from the pattern of citations. No step in that process consults a record of a specific paper.
Two more features make the problem worse:
- A model writing from memory keeps no separate list of the documents it learned from, so it cannot tell which of its references it saw and which it assembled. We explain this in why AI models cannot tell you which of their references they made up.
- Text generation has no built-in pause for uncertainty. A request for references reads to the model as a request to produce references, and fluent text comes out either way.
Some AI tools search a database before answering, an approach called retrieval-augmented generation. It reduces the room for invented references but does not remove the risk, because the model can still misstate what a retrieved source says or attach it to the wrong claim. See what retrieval-augmented generation is, and whether it stops hallucinations.
How common are fabricated citations?
Walters and Wilder (Scientific Reports, 2023) studied 636 citations across 84 literature reviews generated by an earlier version and a newer version of a widely used chatbot. The reviews covered 42 topics from many fields, so the study was not specific to medicine. Of the earlier version's citations, 55 percent were fabricated, compared with 18 percent for the newer version.
In biomedicine, Topaz and colleagues (The Lancet, 2026) report that prior studies estimate 30 to 69 percent of references generated by large language models in biomedical contexts are fabricated. Their own audit of 97.1 million structured references across about 2.5 million biomedical papers found 4,046 fabricated references across 2,810 papers.
Why do fabricated citations matter in health information?
A citation tells the reader that someone checked. In health writing, readers use that signal to decide what to trust, and not every reader opens every source. A fabricated reference borrows the authority of the literature for a claim that may have none.
Once a fake reference appears in a published paper or a widely shared answer, other writers can copy it forward without opening it. Each copy makes the claim look better supported than it is. The Topaz audit found that 98.4 percent of affected papers had received no publisher action at the time of the audit, so fake references can stay in the record.
How do you catch a fabricated citation?
- Look up the PMID or DOI and confirm it opens the paper described.
- Search the exact title on PubMed, the free search engine for biomedical literature maintained by the National Library of Medicine.
- Check that the authors, journal, year and volume match the record.
- Read the abstract against the claim, which also catches misattribution.
The step-by-step version is in how to check if a PubMed citation is real.
A better habit is to reverse the order: find the sources first, then write from them. Luma is built that way. It searches PubMed for each claim in an answer, attaches a source only when the retrieved evidence supports the claim, and flags the rest instead of supplying a citation.
Frequently asked questions
Is a fabricated citation the same as an AI hallucination?
A fabricated citation is one kind of AI hallucination: a confident output with no basis in a real source. Hallucinations also include invented facts and real sources cited for things they do not say.
Why don't AI tools just say they cannot find a source?
Text generation has no built-in step for checking whether a source exists, so a model produces a fluent reference instead of a gap. Saying "not supported" has to be designed into a system on purpose.
Can you ask an AI tool to check its own references?
Asking a model whether its references are real produces more generated text, drawn from the same patterns that produced the reference. Verification has to happen against the source itself, in a database such as PubMed.
Are fabricated citations only an AI problem?
No. People can invent references or copy ones that were already wrong, and fabricated references appear in published biomedical papers. AI tools make the problem easier to create because they can generate a convincing reference on request.
References
- Topaz M, Roguin N, Gupta P, Zhang Z, Peltonen LM. The Lancet. 2026;407(10541):1779-1781. doi:10.1016/S0140-6736(26)00603-3. PMID 42107362. https://pubmed.ncbi.nlm.nih.gov/42107362/
- Walters WH, Wilder EI. Scientific Reports. 2023;13(1):14045. doi:10.1038/s41598-023-41032-5. PMID 37679503. https://pubmed.ncbi.nlm.nih.gov/37679503/