When a court catches an AI making up fake case citations, how is that different from a lawyer's honest typo?

This explores how judges tell a fabricated AI citation apart from an honest clerical mistake, and whether the difference matters to the outcome. The collection has no paper that describes a formal legal test for separating the two. What it does show is more surprising: the courts that have been studied mostly don't treat the line between them as the deciding question. In five cases where courts found fabricated or suspected AI citations, none imposed a penalty for the hallucination itself. Sanctions depended on the judge's discretion, on whether the error caused real harm, and on whether the lawyer acted in bad faith Do courts actually sanction fabricated AI citations when detected?. So in practice, courts handle AI errors the way they already handle careless filings. They ask about intent and consequences, not about which tool produced the mistake.

There's a good reason the 'how did this error happen' question is hard to answer from the outside. One line of work argues that 'hallucination' is the wrong word. A language model produces accurate and inaccurate text through exactly the same statistical process, so its errors are better called fabrication than a misperception or a slip of memory Should we call LLM errors hallucinations or fabrications?. That marks a real difference from a typo. A typo is a slip in recording something real: the case exists, but the page number is wrong. A fabricated citation can point to a case that never existed, written in the same confident, well-formatted style as a real one. Formal proofs suggest no computable model can avoid this completely, which is why outside checks are necessary rather than optional Can any computable LLM truly avoid hallucinating?.

The legal industry's own tools don't remove the problem. A preregistered evaluation found that Lexis+ AI, Westlaw's AI-assisted research and Ask Practical Law AI produced hallucinations in 17% to 33% of answers, despite being marketed as 'hallucination-free'. Because these systems are closed, outsiders can't easily verify where an error came from How often do legal AI tools actually hallucinate citations?. A lawyer who files a bad citation from one of these tools may honestly not know whether they introduced the error or the software did.

The broader research explains why judges catch these errors by checking the citation, not by reading closely. People perform at roughly chance when trying to spot AI-generated content Can people reliably spot content made by AI?. AI text differs from human writing in ways statistics can measure, yet even trained linguists can't perceive those differences Can humans detect AI text if machines can measure it?. In an 81-person study, readers fell for fluent fabrications as readily as for true statements, until an interface showed them which claims were backed by sources Can readers tell truth from fabrication without evidence signals?. Fake citations also work on AI evaluators. LLM 'judges' give higher scores to answers that include fake references, which shows how much citations alone can make a weak argument look credible Can LLM judges be fooled by fake credentials and formatting?.

The takeaway: the useful line isn't 'AI error versus typo' but 'mistake in recording a real source versus a source that doesn't exist'. The only reliable way to tell them apart is to look the source up. Courts seem to have accepted this, focusing on whether lawyers checked their work and whether anyone was harmed. If you're interested in where this goes next, the provenance study suggests the real fix is interfaces that show which claims are backed by sources, not better detectors.