How to Check Whether an AI-Generated Citation Is Real
An AI assistant hands you a reference. It has an author who exists, a journal that exists, a year that fits, a volume and page range that look ordinary, and a title that describes precisely the thing you were arguing. It is also, maybe one time in several, not a real paper.
This is not a bug that got fixed. Language models produce text that is statistically plausible, and a citation string is one of the most patterned objects in academic writing — author, year, title, journal, volume, pages. A model can assemble a flawless-looking one without any of the parts having ever appeared together. Worse, the failure is invisible at a glance: fabricated references are designed, by the nature of the process, to look right.
So don't read citations. Check them. Here is the routine, in the order that kills the most fakes for the least effort.
The five-minute routine
1. Resolve the DOI first
If the citation has a DOI, paste it after https://doi.org/ and open it. A real
DOI redirects to a publisher page. Three outcomes:
- It resolves and the landing page matches the citation — good sign, keep going.
- "DOI not found" — the DOI is invented, or was mistyped. Treat as fake until proven otherwise.
- It resolves, but the article is not the one described — the model attached a real DOI to a fabricated title. This one catches people out, because a resolving link feels like proof.
That last case is why the check is "does the landing page match", not "does the link work".
2. Ask Crossref, not the chatbot
Crossref is the registry behind most academic DOIs, its API is free, and it answers the only question that matters: does this record exist as described?
https://api.crossref.org/works?query.bibliographic=TITLE+HERE&rows=3
Paste that into a browser with the title in place, and read the title,
author, container-title and issued fields of what comes back. If the real
record differs from your citation in year, journal or author list, your citation
is wrong even if the paper is real — and a wrong citation fails a bibliography
check just as hard as an invented one.
If you would rather not read raw JSON: our free citation generator does this lookup for you — paste the DOI, and the fields come back from Crossref, formatted, with nothing invented in between. If the DOI is fake, nothing comes back. That is the answer.
3. Search the title, in quotes, in a real index
No DOI? Then the title is your key. Search it in quotation marks in Google Scholar, PubMed (for biomedicine), or Semantic Scholar. Exact-phrase search matters: without quotes, search engines helpfully find you similar papers, which is precisely the confusion you are trying to resolve.
Zero exact matches across two indexes is strong evidence the paper does not exist. One caveat — very recent preprints and non-English work are indexed unevenly, so for those, check the author's own publication list or their lab page before concluding anything.
4. Check that the paper says what you were told it says
The subtler failure mode is a real paper, misattributed claim. The model cites something genuine, but the finding it credits to that paper is not in it — or is in it with the opposite sign, or with caveats that undo your argument.
Open the paper. Search inside it for the specific number, term or claim. This is the step everyone skips, and it is the one that survives peer review the least well.
5. Check the paper hasn't been retracted
A citation can be real, correctly quoted, and still poison your argument if the paper was retracted. The Retraction Watch database is open and integrated into Crossref data, so retraction status travels with the DOI record. If you are citing something that carries a claim your work depends on, spend the extra thirty seconds.
Red flags worth a second look
| Signal | Why it matters |
|---|---|
| Title reads like a perfect summary of your point | Models generate toward your prompt; real titles are rarely so obliging |
| Round page ranges (e.g. 100–110) | Fabricated numbers cluster on tidy values |
| Author is famous, journal is famous, topic is niche | Plausibility assembled from prestige, not from a record |
| No DOI on a post-2000 journal article | Most have one; its absence is a question, not a verdict |
| The reference list has no duplicates or oddities | Real bibliographies are messy — repeated authors, preprints, chapters |
None of these is proof. They are triage: they tell you which references to check first when you have forty of them and an hour.
What to do when it is fake
Delete it and find the real source for the claim — do not "fix" the citation by adjusting fields until it resolves to something. If the claim was true, a real source exists; if you can't find one, the claim was part of the fabrication, and the sentence has to go too.
Then check the rest of that answer. Fabricated references rarely arrive alone.
The workflow that avoids the problem
The cheapest fix is to stop taking bibliographic data from a model at all. Ask the assistant for ideas, arguments and search terms; get sources from indexes, and citations from the page or record itself, where the metadata is authored by the publisher rather than predicted by a model.
That is the entire premise of ClipCite: when you are on the page you actually read, the citation is built from that page's own metadata — Highwire, Dublin Core, Open Graph tags — and enriched from Crossref when a DOI is present. There is no generation step in the middle to hallucinate through, because the data comes from the source rather than from a prediction about the source.
Verification is still your job. But it is a much shorter job when nothing in the chain invented anything.