Did ChatGPT Make Up This Citation? A 5-Minute Check for a Single Reference

It looks perfect. Plausible authors, a journal you recognize, a year that fits the argument, even a DOI. But something about it feels off — maybe the title is a little too neatly matched to your point, or you can't remember where it came from. If an AI assistant produced that reference, "looks right" tells you almost nothing. Language models are built to generate text that resembles a real citation, not to retrieve one. Here's how to settle the question in about five minutes.

Why AI-generated references look so convincing

A hallucinated citation isn't a random string. The model has seen thousands of real references, so it reproduces the shape of one flawlessly: correct author-name conventions for the field, a journal that genuinely publishes on that topic, a volume and page range in a believable format, and a DOI with a valid-looking prefix. What it can't do is guarantee that the specific combination ever existed.

This is why the usual instincts fail. Formatting checks won't catch it. Neither will "I've heard of that journal." The failure mode is a real-sounding paper that was never written — or, just as often, a real paper stitched to the wrong author, wrong year, or a DOI belonging to something else entirely.

The five-minute check, in order

Work from the most specific identifier to the least. The moment something fails, stop — you already have your answer.

1. Resolve the DOI. Paste it into https://doi.org/ followed by the identifier. A real DOI lands on a publisher page. If you get "DOI Not Found," the reference is almost certainly fabricated. If it resolves but shows a different paper, you have a mismatch, not a fake — the citation needs fixing, not deleting.
2. Search the exact title in quotation marks. Use Google Scholar, Crossref, or PubMed for biomedical work. Quotation marks force an exact-phrase match. A real paper published in an indexed journal will surface. Zero results for an exact title is a strong signal.
3. Search the author plus a keyword. If the title returns nothing, check whether the author exists and works in this area. Sometimes the model has attached a fictional title to a genuine researcher — worth knowing, because that researcher may have written something you can actually cite instead.
4. Check the journal, volume, and year line up. Open the journal's archive for that volume. If volume 34 was published in 2019 and your citation says 2022, the metadata is wrong even if the paper is real.
5. Confirm the paper says what you claim it says. A verified reference that doesn't support your sentence is still a problem — just a quieter one.

What each outcome actually means

It helps to sort results into three buckets rather than a binary real/fake:

Result	What it means	What to do
Verified	The paper exists and the metadata matches	Keep it; confirm it supports your claim
Mismatch	The paper is real, but a field is wrong	Correct the year, authors, or DOI
Unfound	No trace in any index	Remove it and find a real source

The middle category catches people out. Reviewers and supervisors treat a wrong year the same way they treat a wrong claim: as evidence that the bibliography wasn't checked.

When one citation becomes forty

The manual method above is fine for a single suspicious reference. It doesn't scale to a full AI-generated reading list or a bibliography assembled across months of drafting, where you can no longer remember which entries you found yourself and which a chatbot suggested. At that point you want every entry checked against the actual literature databases in one pass.

That's what the Citation Checker at citely.ai does — paste a reference or a whole list, and it tells you which entries are verified, which have metadata that doesn't match, and which don't appear to exist at all. Worth running before you submit, not after someone else finds the one that isn't real.