THE HALLUCINATION MECHANISM
Large language models predict plausible next tokens; they do not retrieve from a database of real cases. A fabricated citation like 'Smith v. Jones, 412 F.3d 891' looks correct because the model learned the SHAPE of citations, not their content. The format is real; the case is not.
RULE 11
Federal Rule of Civil Procedure 11 requires that every filing be signed by an attorney certifying that claims are 'warranted by existing law' after 'reasonable inquiry.' Citing a case you never read — AI-generated or not — has been sanctionable since 1983. The AI era did not create the duty; it created a faster way to violate it.
THE MATA PRECEDENT
In Mata v. Avianca (SDNY, 2023), two attorneys submitted a brief with six fabricated cases generated by ChatGPT, including invented quotations and fake docket numbers. Judge Castel fined them $5,000 and required notification of every judge falsely cited. It became the canonical warning, cited in nearly every subsequent AI-hallucination sanction.
WHY APPELLATE IS WORSE
Trial courts have discovery and live witnesses; errors surface. Appellate courts decide purely on the written record and cited precedent. A fake citation in an appeal brief can shape the legal reasoning a panel applies — and appellate opinions become precedent themselves, meaning a hallucination uncaught at this level can propagate into future case law.
THE VERIFICATION GAP
Westlaw and LexisNexis charge per-search fees that solo practitioners avoid; free AI tools fill the gap. The result: the lawyers most likely to use unverified AI output are those serving clients least able to afford a second opinion. The technology lowers the cost of producing a brief faster than it lowers the cost of checking one.