How to check if an AI citation is real

In short: work through the five steps below in order, and stop at the first failure: a citation that fails step 1 does not get rescued by a good-looking step 4. Each step answers one question — does the identifier resolve, does the record exist, is it the paper claimed, does it support your claim, and is the version right?

An AI-generated citation can look perfect: plausible authors, a real-sounding journal, a plausible year, a well-formed DOI. That is exactly what makes it dangerous. A citation that looks right is not the same as a citation that exists, and a citation that exists is not the same as a citation that says what you claim it says.

This guide answers one question only: is this citation real? It is written so a reader who never contacts Oryx can run the whole check with free tools. The broader pre-submission routine — checking every citation in a paper, not just the suspicious ones — is covered in Oryx’s published guide, “How to check AI-generated citations in your own paper”.

Why “looks real” is a weak test

AI systems generate text by predicting what comes next, not by retrieving a record. A reference list produced that way can mix real papers, real papers with wrong details, and papers that were never written — all in the same confident format. Formatting proves nothing: a DOI-shaped string is not a DOI, and a journal name you recognise is not proof that the paper appeared in it.

Published pilot evidence, labelled as pilot evidence and not a universal rate, shows how often the surface fails: in Oryx’s published pilot audit of 66 AI-generated citations (Paper 2, study 1), 31 citations (47%) had broken links, 7 (11%) mismatched the claim they were attached to, and 11 (17%) could not be verified — only 5 both resolved and fully supported the claim. The point is not that your citations will fail at those rates. It is that a citation you have not checked is an unverified claim, however convincing it looks.

The real-or-fabricated check, in order

Work through these steps in order, and stop at the first failure: a citation that fails step 1 does not get rescued by a good-looking step 4.

1. Does the identifier resolve? Follow the DOI or link. Use a DOI resolver (for example, doi.org) rather than trusting the string’s shape. A DOI that resolves to a different paper, or to nothing at all, is a failed citation — not a formatting problem to tidy up later. If there is no DOI and no link, treat that as a missing identifier, not as a minor omission: it removes the fastest route to verification.

2. Does the bibliographic record exist independently? Search the title, in quotation marks, in an independent bibliographic source — Crossref, OpenAlex, Google Scholar, or Semantic Scholar. You are looking for a record that was created by someone other than the AI system that produced your citation: a publisher page, a journal listing, or a bibliographic database entry.

• Found, matching: move to step 3.
• Found, but with different details: the citation is wrong even if the paper is real. A wrong year, wrong authors, or wrong journal means the reference as written does not describe the paper it points to. Correct it from the record, not from memory.
• Not found anywhere: treat it as fabricated until proven otherwise. A perfect-looking reference that appears in no database, on no publisher site, and in no scholar index has no evidence of existence. Do not keep it on the strength of its appearance.

One caution: absence from one database is not proof of fabrication — some legitimate work (very recent preprints, some conference papers, non-English publications) is indexed slowly or thinly. Absence from every independent source, combined with no publisher page and no author page listing it, is the signal that matters.

3. Is it the paper the citation claims it is? Compare the record against the citation, field by field: title, authors, year, journal or venue, and volume or pages where given. AI-generated citations commonly get one field wrong while the rest look right — a real title attached to the wrong year, a real author list attached to a paper they did not write. Any mismatch means the citation as written is incorrect, even if a real paper with a similar title exists.

4. Does it say what you claim? Finding the paper is only half the check. Read the relevant passage — the abstract at minimum, and the section the claim rests on where the claim is load-bearing. Mark it honestly: supports, partially supports, or does not support. A real paper cited for a claim it does not make is still a failed citation, and it is the failure most likely to survive a quick check, because the reference itself looks legitimate.

5. Is the version right? Check whether the record is a preprint, a conference version, or the peer-reviewed journal version — and cite the version you actually relied on. A preprint cited as peer-reviewed work misstates the evidence’s status, and a journal version cited with preprint details creates a mismatch a reviewer will catch. If the version you found differs from the version cited, correct the citation to the version you verified.

Five signs a citation may be fabricated

None of these alone proves fabrication, but each is a reason to run the full check rather than trust the reference:

1. A DOI that does not resolve, or resolves to an unrelated paper.
2. A title that returns no exact-match result in any bibliographic database or search engine.
3. Authors who do not list the paper on their own publication pages, where such pages exist.
4. A journal and year combination that does not fit — for example, a volume number the journal had not reached in that year.
5. Suspiciously generic detail — a paper whose title, abstract-style claim, and venue all sound right but which no source outside your draft can confirm.

Conversely, do not treat a citation as fabricated merely because it is obscure, old, or in a language you do not read. Obscurity is a reason to check harder, not a verdict.

What to do with the result

• Real and supporting: keep it, exactly as the independent record gives it.
• Real but wrong in detail: correct the citation from the record, and re-check that the corrected paper still supports your claim.
• Real but not supporting: either re-source the claim to a paper that does support it, or drop the claim. Never keep a real-looking citation for a claim it does not support.
• Not found anywhere: drop it. Do not “fix” a fabricated citation by editing its details until it resembles a real paper — that produces a different wrong citation, not a right one. Replace it with a real source you have located and read, or drop the claim.
• Unverifiable after a genuine search: do not cite it. A citation you could not verify is not a citation you can defend to a reviewer.

Keep a one-line log per citation you checked: citation, verdict (real / corrected / dropped), and date checked. If a reviewer asks, that log is your defence — and it is also how you catch the citations you meant to check but did not.

Common traps

• Trusting the DOI’s shape. A string that looks like a DOI is not evidence; only resolution is.
• Checking only the first result. A title search that returns a similar paper is not a match. The record must match the citation, field by field.
• Confusing “the paper exists” with “the paper supports me”. Steps 2 and 4 are separate checks, and passing the first does not pass the second.
• Keeping a citation because removing it leaves a gap. A gap in your references is honest and fixable; a fabricated citation is neither.
• Assuming a real journal name guarantees a real paper. Fabricated citations routinely borrow real journal names.

Sources and further reading

• Paper 2 v2, Trust, But Verify (Zenodo DOI 10.5281/zenodo.23731235) — the published pilot citation audit quoted above.
• Oryx Working Paper v3, From Answers to Auditable Intelligence (Zenodo DOI 10.5281/zenodo.23718472) — the six-layer framework behind the claim-level check.
• Oryx’s published guide, “How to check AI-generated citations in your own paper” (oryxai.in) — the full pre-submission routine this page’s real-or-fabricated check fits inside.

Related: How to check AI-generated citations in your own paper — the full pre-submission routine this check fits inside.

---

Oryx AI offers claim-level evidence audits for research drafts and vendor claims — ₹7,500 per pilot report, payable only on acceptance, currently invite-only beta (contact: amartya@oryxai.in).

Oryx Centre for Strategic Futures is a two-person research initiative (Udyam-registered micro enterprise) working on auditable AI.