The interesting part of this one is not that seven cases turned out to be invented. That happens most weeks now. The interesting part is the sentence the lawyer wrote when she explained how it happened.
She thought the tool was part of her firm’s Westlaw subscription. She thought Westlaw had run the cite check.
What Happened
The case is Meni-Siliga v. A’s Contractor, Inc., in the Superior Court of California, County of Los Angeles. State Farm General Insurance Company is a defendant, represented by outside counsel at Musick, Peeler & Garrett.
On March 31, 2026, that firm filed eight motions in limine on State Farm’s behalf. Motions in limine are the housekeeping of a trial: short, procedural, arguing about what a jury will and will not be allowed to hear. They are the least glamorous filings in a case and among the most heavily precedent-driven, because every one of them cites the evidentiary rule it is asking the judge to apply.
On August 7, 2026, at the Final Status Conference, opposing counsel told the court that the authorities did not check out.
A review found seven case citations across those eight motions that do not exist.
The Declaration
Jacquelene Robinson, a senior associate at the firm and second chair trial counsel for State Farm, filed a declaration shortly after that conference.
She did not blame a paralegal. She did not claim a clerical error. She wrote, plainly: “I regrettably did not check some of the cites I chose to include in the filed pleadings... I should have... neglected to perform a cite check on all of the authorities cited in the motions.”
Then she named the tool. She had used an AI research product called Irys while preparing the motions.
And then she explained why she had not verified its output. She believed Irys was tied to the firm’s existing Westlaw subscription, and that it therefore performed an internal cite check as part of what it produced.
Why That Sentence Matters More Than The Seven Cases
Every AI citation story so far has had the same moral, and it is a moral about discipline. A lawyer pasted from a general purpose chatbot. A lawyer was lazy. A lawyer did not verify. The fix, in every write up, is the same: do not trust the chatbot, run the citator.
That fix does not reach this case.
Robinson was not using a general purpose chatbot. She was using a legal research product, inside a professional workflow, with the belief that it sat behind the same vendor relationship her firm already paid for and trusted. Under that belief, running a separate citator is not diligence. It is redundancy. You do not re-check the thing you believe has already been checked.
The failure was not that she skipped a step she knew existed. It was that she believed the step had already been taken by somebody else.
That is a different problem and it does not get solved by telling lawyers to be more careful.
The Distinction Nobody Is Drawing Clearly
There are now at least three different things a lawyer might mean when they say they used AI for research, and they carry completely different risk:
One. A general purpose model with no retrieval. It generates citations from a language model’s memory of what citations look like. Every output needs to be verified from scratch. Everyone knows this now.
Two. A legal tool with retrieval over a real case database. The citations come from documents that exist. The risk moves from invented cases to mischaracterised ones, which is a smaller and different problem.
Three. A legal tool that looks like the second and behaves like the first. This is the dangerous category, and the reason it is dangerous is that nothing about the interface tells the user which one they are in.
A lawyer who has correctly learned lesson one, and correctly learned that legal-specific tools are safer than consumer chatbots, has been handed exactly the belief that makes category three lethal.
The Part That Is Still Open
The plaintiff asked the court to deny the affected motions and to issue an order to show cause on sanctions, pointing at a 2025 California appellate decision that sanctioned fabricated AI authorities.
As of September 5, 2026, no public reporting identified any sanction imposed on the lawyers or on State Farm over these filings. The motions were filed in March. The problem surfaced in August. Nothing has been decided.
That gap is worth noting on its own. Five months of a trial calendar ran with seven invented authorities sitting in the file, and the thing that found them was not a court, not the firm, and not the client. It was the lawyer on the other side, at the last conference before trial.
What This Should Change
For firms. The question a firm needs to answer is not “do we allow AI research.” It is “does every lawyer here know, for each tool on their desktop, whether its citations have been checked against a real database, and by whom.” Robinson could not answer that question about a tool she was actively using on a client matter, and there is no reason to think she is unusual.
For vendors. If a product outputs a citation that has not been verified against a citator, the product should say so on the citation, every time, in the output itself. Not in the terms of service. Not in an onboarding slide. On the line.
For everyone else. The reassuring version of this story, the one where careless lawyers get caught and careful ones are fine, is finished. The lawyer here told the court exactly what she believed and why, and what she believed was reasonable. It was also wrong, and seven cases that never existed went into a court file for a client that is one of the largest insurers in the United States.
The lesson is not be more careful. It is find out what your tools actually do.