AI Hallucinated Citations: What Attorneys Need to Know

Large language models do not look up citations in databases. They predict the next token in a sequence based on patterns learned during training. When asked to cite a legal authority, the model generates text that looks like a citation because it has seen millions of citations in its training data. But the specific combination of volume, reporter, and page number may not correspond to any real case.

This is not a bug in the traditional sense. The model is doing exactly what it was designed to do: produce plausible-sounding text. The problem is that plausibility and accuracy are different things in law. A citation that looks right but points to a non-existent case is worse than no citation at all.

The hallucination patterns are predictable. Models tend to generate citations with common reporter abbreviations (F.3d, S.W.3d, N.E.2d) because those appear frequently in training data. The volume numbers are usually within realistic ranges. The case names often combine real party names in new combinations. The result is a citation that passes a casual visual inspection.

What makes this dangerous for attorneys is that the output format is correct. The Bluebook formatting is right. The parenthetical with the court and year looks legitimate. An attorney reviewing a draft might scan the citations, see familiar formatting, and assume they were pulled from Westlaw or LexisNexis.

The defense is simple: check every citation against an authoritative database before filing. This is not new advice. Attorneys have always been responsible for the accuracy of their filings. AI tools have simply made it possible to generate a larger volume of citations that need checking.

Certavi was built specifically for this problem. The extraction engine identifies every citation in a brief, and the verification engine checks each one against CourtListener, Westlaw, LexisNexis, or PACER. No AI is involved in the verification decision. The status of each citation traces to a deterministic database query.