[1]

The problem is not the citation. It is the inference.

A wrong page number is a typo. A reference to a paper that was never published is different in kind, because there is only one way it got into your document: something generated it, and nobody checked.

That inference is immediate and it is difficult to argue with. Once a reader knows one citation was invented, every claim in the document becomes a question rather than a statement.

1. Someone checks a single reference

   A guide reading a draft, an examiner preparing for a viva, a reviewer verifying a claim. It takes about twenty seconds.

2. They check the rest

   Nobody who finds one fabricated citation stops there. The whole list gets read differently.

3. The question changes

   It stops being about references. It becomes: which parts of this did you write, and which parts did a model write for you?

4. Process takes over

   At a university that means the thesis goes back. At a journal it can mean an integrity investigation and a retraction, with your name permanently attached to the notice.

[2]

Language models invent citations that look better than real ones.

A model predicts plausible text. A citation is highly patterned text, so a model is very good at producing one that has the right shape — credible authors, a journal that sounds like it should exist, a volume and page range in the right format — without any of it corresponding to a published paper.

Researchers writing in the Annals of the Royal College of Surgeons of England examined more than a thousand references produced by nine chatbots answering surgical questions. In the worst-performing model, around a third of the references were fabricated or could not be verified. Some invented citations carried plausible titles and were attributed to well-known institutions.1

Fabrication is only half of it. A team at Deakin University had GPT-4o write six literature reviews and verified all 176 citations against Google Scholar, Scopus, PubMed and publisher databases. Around one in five was entirely invented. Of those that did point to real papers, 45% carried bibliographic errors, most often an incorrect or invalid DOI. Fewer than half of all citations were both real and accurate.2

That study also found fabrication was worst on the narrow, less-studied topics — around 28–29% on the two lower-profile conditions, against markedly less on the well-known one.2 A thesis topic is, by design, narrow and under-studied. That is precisely where a model has least to draw on, and invents most.

What a fabricated reference usually looks like

Real, findable researchers in the field, placed on a paper they never wrote.

A journal name assembled from words that belong together — Journal of Pediatric Neuroscience Research — that no publisher has ever issued.

A DOI that opens a real paper on a different subject. In the Deakin study most fabricated citations carried a DOI, and around two-thirds of those led to genuine but unrelated articles.2 Clicking the link is not a check.

Perfectly formatted Vancouver punctuation, because formatting is exactly what a model is good at.

You cannot spot these by reading. That is the entire difficulty. They are designed, by the mechanism that produced them, to survive a read.

[3]

It is already in the published record, and it is already being found.

This is not a hypothetical risk being described to make you cautious. Fabricated references have reached print, and screening has caught up with them.

Peer review missing 53 papers should tell you something useful: reviewers are not the safety net. They read for argument and method, not for whether reference 34 exists. The person most likely to check yours is your guide, and the moment they do is the worst possible moment to find out.

[4]

Checking is mechanical. Do it before someone else does.

Every real journal article has a record in Crossref or PubMed. Either a reference resolves to one or it does not, and that question has a definite answer that takes a database lookup rather than a judgement.

Doing it by hand means searching each reference, comparing the title, author, year, volume and pages, and repeating that eighty times without losing concentration. Most people start carefully and stop being careful somewhere around reference twenty.

Books, chapters, government reports and theses are a separate matter — these databases do not index them, so their absence means nothing. A check worth trusting has to know the difference between a reference that is missing and a reference that was never going to be there.

Our checker reads your reference list, looks up every entry, and tells you which ones do not exist. See the count for free before deciding whether you need the detail.