Skip to main content
Citation Safe

What is an AI-hallucinated citation?

https://citationsafe.com/faq/what-is-an-ai-hallucinated-citation

An AI-hallucinated citation is a citation a generative AI tool produced that does not hold up against a primary source — because the case or statute it names doesn’t exist, because a quotation attributed to a real source never appears in it, or because a real, correctly-cited case doesn’t actually stand for the proposition it’s attached to. The term comes from the broader “AI hallucination” problem: large language models generate the statistically plausible next token, not a fact they looked up, so a citation can read exactly like a real one — correct-looking reporter volume, plausible court and year — while referring to nothing.

The three failure modes

In practice, an AI-hallucinated citation fails in one of three distinct ways, and it matters which one because each requires a different check to catch:

Existence-only citation checkers catch the first failure mode reliably. They generally do not catch the second or third, which is why a citation can pass a basic “does this case exist” check and still be functionally fabricated in the way that matters to a judge.

Real examples

Mata is the existence-failure archetype: six ChatGPT-invented cases, none real. Park v. Kim is a single existence failure in a reply brief that the Second Circuit itself caught while preparing for oral argument — proof that a hallucinated citation doesn’t need to be caught by opposing counsel to end in sanctions.

Why AI tools produce these with such confidence

A generative AI model is not a database lookup. When asked for supporting authority, it produces text that pattern-matches the surface texture of a real citation — plausible party names, a correctly formatted reporter citation, a court and year that fit the fact pattern — because that is what its training optimized it to do. It has no internal notion of “I don’t actually know this case,” and when asked to confirm a citation is real, it will typically just generate another confident-sounding confirmation, which is exactly what happened in the Mata record. That is why verifying a citation by re-asking the same model, or a different model, does not work — see how to check if a case citation exists for what does.

Related resources

Verify a citation now, or get sanctions in your inbox weekly

Run a free check — 3 verifications a month, no card required. Or subscribe to AI-Citation Sanction Weekly for one email a week pulled from this same database.

Written by the Citation Safe Research Desk · Reviewed by Andy Gaber, Founder