How to check whether a citation is real

AI tools now produce references that look flawless and don't exist — the right author list, a realistic title, a real journal, a well-formed DOI, for a paper nobody wrote. Formatting is the easy part to fake. Here is a fast, reliable way to tell a real reference from a fabricated one, whether it came from a chatbot, a co-author, or your own notes.

The five checks, fastest first

Run these in order and stop as soon as one fails. Most fabricated references die at step one.

1. Resolve the DOI. Paste the DOI into doi.org. A real DOI lands on the exact paper. A 404, or a redirect to an unrelated article, means the reference is almost certainly fabricated. No DOI at all is a reason to look harder, not to relax.
2. Search the exact title. Put the full title in quotation marks and search PubMed (for biomedical work) or Crossref. A real paper turns up with the correct metadata; a fabricated one returns nothing, or a real paper with a different title.
3. Match author, journal, and year. Confirm all three describe the same real paper. Fabrications often pair a real author with a journal they never published in, or attach a plausible year to a volume that doesn't line up. Any mismatch is a red flag.
4. Check for a retraction. PubMed flags retractions prominently, and Retraction Watch maintains a database. A real but retracted paper is not a citation you want to rest a clinical claim on.
5. Open the source and read it. The final step, and the one people skip: open the paper and confirm it actually says what your sentence claims. A reference can be real and still not support the specific point attached to it.

A reference you can't open is a reference you can't defend. If you've exhausted every check and still can't find the paper, remove it and re-source the claim.

The three kinds of fake citation

Fabrications aren't all the same, and knowing the shape helps you catch them.

Why this is worth the minutes

This isn't a hypothetical risk. In a JAMA Ophthalmology test, roughly a third of AI-generated references couldn't be verified; in one analysis of AI-generated medical content, 47% of references were fabricated outright. Fabricated citations have reached published papers, survived peer review at major conferences, and drawn court sanctions. We keep a running record of the documented cases in the Fabrication Case File.

The limit of checking after the fact

Manual verification works, but it's slow, and it only catches what you have time to check. Automated reference checkers help, yet they still run after the model has already written — they reduce the fabrication rate without closing the gap. The only way to make a fabricated citation impossible is to remove the chance to invent one: retrieve real papers first, and let the model cite only from that retrieved set.

That is the design behind Cento. Every reference is drawn from a candidate set retrieved from PubMed, OpenAlex, Semantic Scholar, and Europe PMC before a word is written; the model can point only at papers that were actually retrieved, and each citation is validated server-side before it reaches you. There is no step at which a paper that doesn't exist can enter. See how Cento makes a fabricated citation impossible.