A Los Angeles Superior Court judge fined a State Farm defense attorney $999.99 on September 15, 2026, after discovering that eight court filings in a homeowner's fire-and-water-damage lawsuit contained seven fabricated case citations, invented quotes, and legal holdings that simply don't exist. The number itself is the tell: Judge Elizabeth Bradley set the sanction one cent below the $1,000 threshold that would have triggered a mandatory report to the State Bar of California.
This is at least the dozenth publicly reported case this year of AI-hallucinated citations reaching a judge — explainx.ai has previously covered the broader pattern, including a public tracker now documenting more than 1,500 such incidents worldwide. What makes this one worth a closer look is exactly how the mistake happened — a detail that should worry anyone relying on AI research tools inside a professional workflow, legal or otherwise.
TL;DR
| Question | Answer |
|---|---|
| Who was sanctioned? | Attorney Jacquelene Robinson, outside counsel for State Farm |
| What happened? | Eight court filings contained seven fabricated case citations from an AI research tool |
| How much was the fine? | $999.99 — one cent under the mandatory State Bar reporting threshold |
| What tool was involved? | An AI legal research tool called Irys |
| What went wrong specifically? | Robinson mistakenly believed Irys was integrated with her firm's Westlaw subscription — so she treated its output as pre-verified |
| What case is this in? | Fa'alagilagi Meni-Siliga v. State Farm, over fire and water damage to a Carson, CA home |
| Is this a one-off? | No — part of a documented pattern of 1,500+ similar incidents tracked worldwide |
What actually went wrong
The specific failure mode here is more instructive than the raw fact of "AI hallucinated some citations." Robinson's stated defense wasn't that she trusted an AI tool blindly — it's that she believed Irys was connected to Westlaw, the industry-standard legal database whose case citations are, by design, real and verifiable. If that belief were correct, treating its output as reliable would have been reasonable. It wasn't correct, and the seven fabricated citations, invented quotes, and non-existent legal holdings went into eight separate filings before anyone caught them.
This is a tool-attribution failure, not just a hallucination failure: the underlying AI model generated fabricated content (which is expected behavior for any generative model asked a question outside its verified grounding), but the actual professional failure was misunderstanding which system she was using and what guarantees it did or didn't carry. That distinction matters for anyone integrating AI research tools into a workflow with legal, medical, financial, or other high-stakes consequences — the question isn't just "is this tool accurate," it's "do I actually know what this tool is querying, and does it tell me when it's uncertain."
That's precisely the gap TypeSafe AI's Jev, launched the same week, is trying to close with calibrated confidence scores attached to every output — though as critics were quick to point out during that launch, a confidence score attached to a wrong answer doesn't make the answer right, it just makes the uncertainty visible if you check it.
The $999.99 detail is not a coincidence
California Business and Professions Code requires courts to report attorney sanctions of $1,000 or more to the State Bar, which can trigger a formal disciplinary inquiry separate from the sanction itself. Setting the fine at $999.99 — one cent under that line — is a judicial choice, not an accident, and it signals the court weighed the misconduct as serious enough to sanction but not (on this occasion) serious enough to escalate to bar discipline. Whether that calibration holds if the same attorney or firm repeats the mistake is an open question; courts across the growing body of AI-hallucination cases have generally shown far less patience for a second offense than a first one.
Part of a fast-growing, well-documented pattern
This case doesn't exist in isolation. A public tracker cited in explainx.ai's broader analysis of AI hallucination legal cases now documents more than 1,500 court filings worldwide where AI-fabricated citations reached a judge — a number that has grown fast enough that "I didn't realize the AI tool could invent case law" is an increasingly hard defense to sustain in front of a court that has seen the pattern dozens of times before. That earlier analysis found something specific and worth restating here: courts consistently punish the cover-up — failing to correct or disclose the fabrication once discovered — far more harshly than the original hallucinated citation. Robinson's filing corrected the record once the fabrications surfaced, which is likely part of why the sanction landed where it did rather than at a more severe tier.
The underlying case, briefly
The lawsuit itself — Fa'alagilagi Meni-Siliga v. State Farm — is a coverage dispute over fire and water damage that left a Carson, California home uninhabitable. That context matters for reading the sanction correctly: the fabricated citations didn't determine the outcome of the underlying insurance dispute, but they did compromise the integrity of the filings supporting State Farm's defense, which is exactly the harm sanctions rules are designed to address regardless of who would have won the case on the merits. Courts generally treat citation fabrication as a standalone violation of an attorney's duty of candor to the court — separate from, and often more consequential than, whatever the fabricated citation was originally meant to support.
Why in-house legal and compliance teams should care beyond law firms
This case involved outside counsel, but the exposure runs through the client too. State Farm didn't generate the fake citations, but its case was the one built on top of them, and the reputational and procedural fallout (a sanctioned filing on the public record, tied to the company's name) lands on the client regardless of who's ultimately liable for the fine. Any organization that relies on outside counsel, contractors, or vendors using AI tools inherently inherits some of that verification risk — which is why a growing number of corporate legal departments have started requiring outside firms to disclose which AI research tools they use and confirm a human-verification step before filing, rather than trusting firm-level AI policies they can't directly audit.
What this means if you use AI for research or drafting
- Verify the source, not just the content. If a tool's output includes citations, case law, or factual claims, confirm independently whether that tool is actually grounded in a verified database (like Westlaw or PACER) or whether it's a general-purpose or loosely-integrated AI system generating plausible-sounding but unverified text.
- Never file, publish, or ship AI-generated citations, statistics, or quotes without checking each one against a primary source. This applies well beyond law — the same failure mode shows up in journalism, academic writing, and technical documentation whenever a generative model is asked to produce specific facts it wasn't explicitly grounded on.
- A confidence score or "verified" label is not a substitute for checking. As the Jev launch discussion made clear this same week, calibration and confidence signals reduce risk, but they don't eliminate the need for human verification on anything with real consequences attached.
What courts are increasingly requiring
As the volume of AI-hallucination sanctions has grown, more courts have started requiring an explicit certification in filings — a statement, sometimes under penalty of perjury, that the attorney has personally verified every citation against a primary source, whether or not AI tools were used in drafting. Some federal districts and individual judges have gone further, adopting standing orders that require disclosure any time generative AI contributed to a filing. Robinson's case, and the broader pattern it fits into, is part of why that trend is accelerating: courts have concluded that trusting firm-level policies without a specific, filing-level certification isn't sufficient anymore, given how often the failure mode repeats even at firms that already had an AI-use policy on paper.
FAQ
What happened with the State Farm lawyer and AI-generated citations? Attorney Jacquelene Robinson filed eight briefs in a Los Angeles case containing seven fabricated citations, invented quotes, and non-existent legal holdings generated by an AI tool she mistakenly believed was integrated with Westlaw.
How much was the fine, and why that specific amount? $999.99 — one cent below the $1,000 threshold that would trigger a mandatory report to the State Bar of California.
What is Irys, the AI tool involved? An AI legal research tool that Robinson mistakenly believed was connected to Westlaw's verified case database, rather than an independently-generating system.
What case was this filed in? Fa'alagilagi Meni-Siliga v. State Farm, a coverage dispute over fire and water damage to a Carson, California home.
Is this an isolated incident? No — a public tracker documents more than 1,500 similar cases worldwide, a pattern this blog has covered in depth previously.
Does this mean lawyers shouldn't use AI research tools? No — the recurring failure isn't AI use itself, it's filing AI output without independently verifying every citation against a real, checkable source first.
Related reading
- AI hallucination legal cases: why lawyers keep getting sanctioned
- Why do AI models hallucinate, and how to catch it
- AI and the law: legal help and contracts guide
- TypeSafe AI's Jev: calibrated confidence, not correctness
- Structured output and JSON mode prompting: a complete guide
- Official: Los Angeles Superior Court · State Bar of California
Details in this piece reflect court filings and reporting available as of September 16, 2026. Sanction outcomes and any related State Bar proceedings may change as the underlying case continues.