How to Verify an AI Citation Before You Use It

An AI citation can fail in four different ways. The source may not exist. The details may belong to a different source. The real source may not support the sentence attached to it. Or the source may have been corrected, retracted or overtaken by later evidence.

Verification therefore has four gates: identity, access, support and status. Stop at the first failed gate. A plausible title, a working DOI and a real journal are clues—not permission to repeat the claim.

1. Freeze exactly what the model supplied

Copy the complete citation and the sentence it is supposed to support before searching. Preserve author names, title, journal or publisher, year, volume, issue, pages, DOI, URL and any quoted wording. Do not silently repair the citation yet. The mismatch is evidence about reliability.

Mark which details came from the model and which came from you. If the answer provides only a title-like phrase, treat it as a search lead rather than a citation. If the model says it cannot verify the reference, keep that uncertainty instead of prompting it to guess a cleaner-looking version.

2. Resolve the identifier, but do not stop there

For a DOI, open it through doi.org or retrieve its metadata through Crossref. Crossref's public metadata can include title, authors, publication container, dates, references, funders and post-publication updates when members deposit them. A DOI that resolves establishes a registered object and its current landing page. It does not prove every citation field is correct or that the article supports the claim.

For biomedical literature, search PubMed by exact title, author and year. PubMed records use a PMID and may link to publisher text or PubMed Central. For books, reports and conference papers, search the named publisher, library catalog or institutional repository. An arXiv identifier should resolve on arXiv and show a version history.

If no identifier was supplied, search the title in quotation marks, then pair distinctive title words with the first author. Crossref's Simple Text Query can match a formatted reference to candidate DOIs. Do not manufacture an identifier from the first similar result.

3. Compare every identity field

Match title, authors, year, venue and identifier as a set. Small punctuation differences are normal. A different population, subtitle, year or first author can mean the model blended two papers. Check whether the date shown is online publication, issue publication or a later version.

When one field conflicts, use the publisher record and registration metadata to determine the authoritative identity. Record the correction rather than presenting the AI's original citation as verified. A citation that mixes a real DOI with an invented title has failed even though the link works.

4. Open the source, not only the search result

Read the abstract and the relevant section of the full text when available. Search snippets, knowledge panels and AI summaries can repeat the same error you are trying to check. If the full text is inaccessible, say what you could verify and what remains unavailable. Do not claim support from a title alone.

Identify the source type. A protocol describes a planned study, not its results. A preprint may not have completed peer review. An editorial or narrative review differs from a systematic review. A conference abstract can lack methods needed to judge the conclusion. These labels do not make a source worthless; they change what it can carry.

5. Test the claim at the right strength

Rewrite the model's sentence as a checkable proposition. Then locate the passage, table or result that bears on it. Compare population, intervention or exposure, comparator, outcome and timeframe. A paper about college students cannot automatically support a claim about children. An association does not establish causation. A laboratory measure does not automatically prove a meaningful real-world benefit.

Use three labels in your notes: supports, partly supports or does not support. Add a short reason. “Partly supports—same direction, different population” is more useful than a green check mark.

6. Verify quotations character by character

Search the source for a distinctive phrase. Confirm the words, speaker or author, location and surrounding context. AI systems can convert paraphrases into quotation marks, combine adjacent sentences or attach real words to the wrong document. If you cannot find the exact wording in the authoritative text, remove the quotation marks and either write a supported paraphrase or omit it.

7. Check corrections, retractions and versions

Look at the publisher page for correction, expression-of-concern and retraction notices. Crossmark can expose current document status when a publisher participates. Crossref also makes Retraction Watch data publicly available through its REST API; the database is updated each working day, although coverage of corrections and expressions of concern is not complete.

For preprints and repositories, inspect version history. A later version may change the result, authorship or wording. A paper can remain useful after a narrow correction, while a retraction can occur for reasons that affect only some claims. Read the notice rather than treating status as a binary badge.

8. Find the evidence around the paper

One real paper can still be an outlier. Search for systematic reviews, replications, later citations and credible contrary results. Use the scientific-paper reading guide to inspect design and uncertainty, then use the language of uncertainty to report what remains open.

For a low-stakes background sentence, identity and direct support may be enough. For medical, legal, financial or safety claims, use current professional or institutional guidance and qualified advice. A verified citation does not turn a general AI answer into individualized guidance.

9. Leave an audit trail

Save the verified citation, stable identifier, access date and the exact passage or table used. Note any version or status notice. If the source is paywalled, record that you verified metadata only. If a colleague must repeat the check, they should be able to follow the path without asking the model what it meant.

Failure conditions

• No matching record: treat the citation as unverified, not “probably real.”
• Blended details: rebuild the citation from the real source and disclose the mismatch.
• Real source, wrong claim: do not cite it for that sentence.
• Abstract only: limit your statement to what the accessible record supports.
• Correction or retraction: read the notice and reassess the affected claim.
• High stakes: verify through current authoritative guidance and a qualified human.

Authoritative verification tools

• Crossref, Metadata Retrieval documentation and REST API.
• Crossref, Retraction Watch data in Crossref.
• Crossref, Simple Text Query.
• U.S. National Library of Medicine, PubMed Help.