State Farm Lawyers Admit AI-Fabricated Case Citations

Reason published excerpts from a declaration by State Farm attorney Kenneth Katel. Katel wrote that, at an August 7 final-status conference, Meni-Siliga's counsel told the court they had found citations in State Farm's motions in limine that were "fabricated" and possibly "hallucinations."

According to the declaration, the first five motions in limine at issue had been filed and served on March 31, 2026. Katel wrote that the allegation was the first time he became aware of the problem in the earlier filings.

Katel's declaration included an apology to the court, Meni-Siliga and her attorneys. He also described a review of State Farm's motions in limine and supporting briefs, oppositions to the plaintiff's motions in limine, and filings connected to summary judgment and reconsideration. He wrote: "Based on that review, I can represent that there are no AI hallucinations within any of State Farm's Oppositions, Replies or the MSJ documents."

HSJ Chronicle reported that the Los Angeles firm Musick, Peeler & Garrett admitted AI tools had been used in preparing filings containing fabricated citations, invented quotations and legal holdings. The available reporting does not establish which model or legal-research product was involved.

State Farm spokesman Tom Hartmann told the Los Angeles Times: "State Farm expects its external counsel to conform to the highest level of ethical standards and professionalism, including in their use of AI and confirming the accuracy of all legal filings."

The reported episode concerns filings by State Farm's outside counsel. Nonetheless, it adds to a growing record of court disputes involving generative AI outputs presented as legal research. In comparable legal workflows, the key technical failure is not merely inaccurate text generation: a fabricated citation can appear plausible enough to bypass a superficial review, while its existence, quoted language and procedural relevance can be independently checked in authoritative case-law databases.

For ML and legal-tech practitioners, the case illustrates why retrieval, citation validation and human review are distinct controls. A system can produce fluent legal prose even when its source grounding fails. Courts and opposing counsel can verify the resulting claims directly, making provenance and document-level verification more consequential than stylistic output quality.