A reference supplied by an AI assistant should only be treated as trustworthy after two separate checks: first, confirm that the work actually exists and that its bibliographic details match; second, confirm that the work supports the claim made by the AI. A valid DOI alone does not prove that the title, authors, edition, or attributed conclusion are correct.
A five-step verification protocol
1. Preserve the reference exactly as the AI supplied it
Before searching, copy the entire reference into a note: title, authors, year, journal or publisher, volume, issue, pages, edition, DOI, URL, and any passage the assistant presented as evidence. This prevents you from mentally correcting the citation while searching and gives you a fixed record to compare with what you find.
2. Test whether the work exists
Start with the most distinctive parts of the citation. For articles, search the title in quotation marks or combine the first author's surname, year, and part of the title. Crossref maintains bibliographic metadata deposited by publishers and other trusted sources and allows records to be queried by title, authors, journal, year, and DOI. Crossref's reference-matching tool can also take a plain-text citation and suggest a corresponding DOI.
For biomedical literature, PubMed is a useful second check. Its own guide recommends pasting the title or entering the author, journal, and publication year to find a specific citation. If no appropriate database can locate the work, that is a warning sign, but not by itself proof that the reference was fabricated: older books, local works, reports, and publications outside those databases may require a library catalog, the publisher's website, or another specialized index.
3. Verify the DOI and compare the metadata
If the AI supplied a DOI, do not stop at checking whether the address opens. Resolve the identifier at https://doi.org/ and compare the resulting record with the citation you received. Check at least the title, authors, year, journal or publisher, and publication type. Crossref provides an endpoint for retrieving the record associated with a DOI registered in its database, and the DOI system is designed to persistently identify a registered digital object or other resource.
There are three important outcomes. First: the DOI does not resolve, which may indicate a typing error, a nonexistent identifier, or a temporary problem. Second: the DOI resolves but points to a different work, which is an objective mismatch. Third: the DOI resolves to the correct work, but one or more fields in the AI-generated reference are wrong. In that case, the reference is not completely fabricated, but it was presented inaccurately.
4. Check the edition, version, and cited object
A reference can look correct and still point to the wrong edition. This is especially relevant for books, manuals, standards, guidelines, reports, datasets, and software. Compare the year, edition number, publisher, authors or responsible institutions and, when available, ISBN, DOI, or version identifier. If the claim depends on a specific edition, open that exact edition. A later edition may have changed tables, recommendations, chapters, or pagination.
When the bibliographic record and the document you can access do not appear to match, look for the official page of the publisher, journal, responsible institution, or repository hosting the version of record. Do not treat a loose copy, a secondary citation, or a search result as an automatic substitute for the stated source.
5. Test the claim inside the source
The decisive step is to verify the statement attributed to the reference. Open the abstract and, when possible, the full text. Search for the central terms in the claim and read the surrounding context. Ask whether the document actually says that, in what population or setting, under what conditions, and with what degree of certainty.
An AI system can cite a real work to support a claim that the work does not make. It can also turn an association into causation, generalize a narrow result, omit a limitation, or attribute to the article a conclusion found only in a secondary source. That is why checking whether a reference exists and checking whether it supports a claim are different tasks.
How to classify the result
| Result | What you found | How to interpret it |
|---|---|---|
| Verified | The work exists, the metadata match, and the content supports the claim in the proper context. | The reference can be used while preserving the source's limitations. |
| Real reference, wrong details | The work exists, but the author, year, title, journal, pages, edition, or DOI is wrong. | Correct the citation before using it. |
| DOI mismatch | The DOI exists but belongs to a different work. | The presented reference is bibliographically incorrect. |
| Real source, unsupported claim | The work exists and the citation details match, but the text does not support the attributed claim. | Do not use that source as evidence for the claim. |
| Not located | Reasonable searches in suitable databases, the publisher, journal, or catalog do not find the work. | Treat it as unconfirmed until verifiable evidence appears. |
Typical signs of a possibly fabricated reference
- a highly plausible title with no record in suitable bibliographic databases;
- an author, journal, and year combination that does not appear in the publication's history;
- a DOI that looks correctly formatted but does not resolve;
- a DOI that resolves to a completely different title;
- volume, issue, or page numbers that do not fit the journal and year;
- a cited edition with no record at the publisher or in a catalog;
- a passage attributed to the source that does not appear in the available abstract or full text.
A quick everyday check
- Copy the reference exactly as the AI gave it.
- Search the title, author, and year in an appropriate database.
- If there is a DOI, resolve it and compare the metadata.
- Confirm the edition or version when relevant.
- Open the source and verify that it actually supports the claim.
- Record the result as verified, incorrect, mismatched, or unconfirmed.
The core principle is simple: a reference is not validated because it looks academic, has a DOI, or is formatted correctly. Verification is only complete when bibliographic identity and claim support have been confirmed separately.
