AI citations: common failure modes
AI-generated bibliographies often look plausible at a glance. The failure modes are usually subtle: a DOI that almost looks right, a real author list paired with the wrong title, or a valid journal name attached to a non-existent paper.
Failure mode 1: Non-verifiable references
The most direct failure: the cited work can’t be resolved in public bibliographic sources using the provided identifier or citation string.
- Made-up DOI patterns (realistic prefix/suffix shape, but not registered).
- PMIDs that don’t exist, or correspond to a different topic.
- Book citations with a plausible title but no matching ISBN/edition.
Failure mode 2: Mismatched metadata
Sometimes the identifier resolves, but the metadata doesn’t match the reference you’re about to ship.
- Wrong year (online-first vs print year, or simply wrong).
- Wrong title paired with a real journal and plausible authors.
- “Citation drift” where the cited paper exists but doesn’t support the claim.
Failure mode 3: Secondary citations and overclaiming
AI can cite a real paper but misrepresent what it proves. This is common when a claim is a paraphrase of a paper’s discussion section, or when the true evidence is in a referenced work (“paper A cites paper B”).
A practical review workflow
- Extract identifiers (DOI/PMID/ISBN) if present; otherwise treat as higher risk.
- Run a batch pass in Citation Verification.
- For any “needs review” item: open the publisher record and compare title/authors/year.
- For the most important claims: read the relevant section and verify it supports the claim.
- If you need formatted output after verification, generate citations via Citation Generator.
Related study
We publish a reproducible benchmark measuring how often citations in AI outputs are verifiable: Citation Verifiability in AI Outputs (Jan 2026).
Next steps
Use Citation Verification to triage identifiers quickly, and Bibliography Health Check to scan full reference lists.