AI Hallucinated Citations: A Researcher Verification Protocol

Treat every AI citation as unverified until it passes two checks

The first check is existence: can you resolve the URL, DOI, PMID, arXiv ID, publisher page, database record, or library item? The second check is claim support: does the source actually say what the sentence needs it to say?

Researchers often stop after the existence check. That misses the most common high-friction failure in AI-assisted writing: a real source is attached to an overstated, outdated, or unrelated claim. Verification has to operate at the sentence level.

• Existence check: source can be opened or resolved in a trusted system.
• Metadata check: title, authors, venue, year, and identifier match.
• Support check: passage, table, figure, or section supports the exact claim.

Classify the failure before fixing it

A fabricated citation needs deletion or replacement. A mismatched citation needs a source that actually supports the claim, or the claim needs to be rewritten. A weak citation may be usable only if the evidence level is disclosed, such as abstract-only or metadata-only.

The fastest way to avoid false confidence is to keep a failure label in the notebook. Use labels such as nonexistent source, bad identifier, metadata mismatch, source does not support claim, abstract-only support, retracted or corrected source, and contradiction found.

Use registries and libraries for the mechanical checks

Crossref is useful for DOI metadata, PubMed for biomedical records and PMIDs, arXiv for preprint identifiers, OpenAlex for broad discovery metadata, and Zotero for preserving the final library state. Claude can help format checks, but these systems should carry the authoritative record.

If a source has no stable identifier, search by title and author in the field database or publisher site. If you still cannot find it, do not cite it as real. Ask Claude to mark the source as unresolved rather than producing a polished replacement citation.

Record the human decision

The final verification artifact should be short: citation, claim, identifier, source status, supporting passage or section marker, evidence level, contradiction status, and human decision. This note can live in Zotero, a spreadsheet, a manuscript comment, or a project notebook.

When a citation fails, preserve the failure. A correction history is valuable because it shows which prompts, source formats, or workflows are producing risk. Over time, the lab learns where Claude is reliable and where it needs tighter guardrails.

Workflow checklist

1. Copy the exact AI-generated citation and the claim it supports.
2. Resolve the identifier or source record in a trusted system.
3. Compare title, authors, year, venue, and source type.
4. Inspect the passage, table, figure, or section tied to the claim.
5. Check retraction, correction, and contradiction signals where relevant.
6. Rewrite, replace, downgrade, or delete the claim based on the evidence.
7. Save a verification note with the final decision.

Researcher FAQ