How to Detect Fake Citations and AI-Generated References

What people mean by a fake citation

A fake citation is not always a completely invented paper. It can be a reference that does not exist, a real DOI attached to the wrong title, or a record assembled from the authors, journal, year, and pages of different works. AI-generated references can look polished enough to pass a quick formatting check.

Use the term carefully: a metadata mismatch is evidence that the reference needs review, not proof that an author deliberately fabricated it. A paper can also exist while a particular citation fails to identify it correctly.

Check identity before judging the claim

Start by preserving the exact reference and separating citation identity from claim relevance. Resolve the DOI when one is present, then compare the returned title, authors, venue, year, pages, and other identifiers. When there is no DOI, combine the strongest available bibliographic signals and inspect multiple candidate records.

A citation checker can establish whether metadata sources agree on the identity of a scholarly work. It cannot prove that the paper supports a sentence in a manuscript, that the research is sound, or that a provider record is complete.

• Preserve the original reference before normalizing or correcting it.
• Treat a DOI that resolves to another work as a high-priority conflict.
• Compare title, authors, venue, year, volume, issue, and pages instead of trusting one field.
• Keep missing data, provider outages, ambiguity, and contradictions as different outcomes.
• Escalate close candidates and high-impact decisions for human review.

Common signals of a hallucinated or mismatched reference

Several signals should raise the review priority: a DOI that resolves to a different title, an author list that never appears with the claimed work, a journal and year combination that conflicts with the volume or pages, or a title that appears only in the AI output and not in scholarly records. None of these signals alone proves fabrication, but each warrants evidence gathering.

Do not fill gaps by combining fields from the top results. That creates a new citation that may never have existed. Keep competing candidates visible and record which source supports each field.

A reviewable fake-citation check

A useful report shows the original input, the selected or proposed work, field-level evidence, candidate matches, provider diagnostics, and unresolved conflicts. A confidence score can prioritize a queue, but it should not hide why a result is uncertain or turn a not-found result into a definitive accusation.

For high-risk references, open the publisher page, repository record, or original publication after the metadata check. This final step is especially important when a citation is central to a claim, a systematic review, or an editorial decision.

Where CiteWise fits

CiteWise is an evidence-backed citation verification workflow for detecting identity mismatches in ordinary and AI-generated references. It compares bounded scholarly metadata sources, recovers a missing DOI when the candidate is supported, and keeps evidence separate from correction proposals. It helps researchers and editors find references that need attention; it does not label a person or paper as fraudulent and does not replace scholarly judgment.