To verify an AI citation, check three separate things: does the work exist, is this the relevant version, and does it support the claim? A working link or a real DOI answers only part of that question. A perfectly real paper can be attached to a statement it never makes.
Check the source's identity
Open the publisher or institution's record and compare the title, authors, year and identifier with the citation. Crossref metadata can help identify many scholarly works, but absence from Crossref is not proof that a source is fabricated: other registration agencies and non-DOI publications exist. A DOI is an identifier, not a truth certificate or a guarantee of peer review.
Do not trust a familiar-looking domain or a title paraphrased by the model. If you cannot locate the exact item, mark it unresolved and ask for a verifiable source rather than guessing which paper the answer intended.
Check version and status
Is the file a preprint, accepted manuscript, final publication or later correction? Does the publisher display an update, expression of concern or retraction? Crossmark can report changes where a publisher participates and registers them. A missing Crossmark button does not establish that no update exists; inspect the publisher record and relevant notices too.
Match a claim to inspected evidence
| Field | What to record |
|---|---|
| Claim | The exact factual statement you plan to rely on. |
| Source identity | Title, author/institution, date and identifier/URL. |
| Inspected evidence | Section, table, figure or passage actually read. |
| Scope | Population, geography, measurement period and conditions. |
| Verdict | Supported, partly supported, contradicted or unresolved. |
| Limit | Access restrictions, missing methods or inferential steps. |
Record only what you inspected. A search-result snippet or abstract is not the complete paper. A source saying two quantities were associated does not automatically support an AI answer saying one caused the other.
An original claim-matching example
Suppose an answer says “Product X cuts research time by 40%.” The linked source is a vendor's hypothetical scenario using four hours before and 2.4 hours after. The arithmetic is a 40% reduction, but that is not an independently measured user outcome. A defensible revision would label it as the vendor's illustration and state the assumptions, or remove it if it does not help the reader.
The example is invented to show the distinction. It makes no performance claim about a real product. The same check applies to customer counts, rankings, clinical outcomes and quoted financial savings: identify both the number and what it measures.
Use AI to organize checks, not certify itself
- Extract a short list of consequential or surprising claims.
- Verify identity and inspect primary material for those claims.
- Revise the wording to match the actual evidence and limitations.
- Keep a dated correction trail for substantive changes.
An AI saying it “verified” its own answer does not provide independent evidence. For a useful distinction when choosing a source, read primary versus secondary sources. For access to scholarly material, see authorized reading routes.