Seven case citations spread across eight court filings pointed at decisions that were never handed down. Some of the case titles were wrong. Some of the quotations appeared in no opinion anywhere. The filings were made on behalf of State Farm, in a Los Angeles Superior Court case brought by a postal carrier who lost her house.
Nobody on the State Farm side caught any of it. The person who caught it was the lawyer on the other side of the table, reading the motions filed against his client and counting.
The House Came First
Fa'alagilagi Meni-Siliga is a United States Postal Service carrier. In 2020 a fire in her home in Carson, in Los Angeles County, left the house uninhabitable. A contractor started the repairs. Then a storm arrived and added water damage on top of the fire damage. She has described a stretch of waiting for State Farm to approve the work, and the work was never finished.
She and her husband drained their savings. They drained their retirement accounts. They declared bankruptcy. The house went to foreclosure. In July 2024 she sued, naming the contractor, the contractor's insurance company, a public adjuster, and State Farm, her own homeowners insurer.
"My family and I have lost forever our family home," she wrote in a court filing. "We do not come from money. It took many years and sacrifice for our family to purchase our home."
State Farm's original answer argued that negligence by Meni-Siliga, her contractor and her public adjuster caused the damages she suffered, and that she had failed to bring the action within a year of the original loss. Trial is set for October.
That is the case. The AI part sits on top of it, and it is worth remembering what is underneath, because the fabricated citations were not filed into an abstract dispute about doctrine. They were filed into a fight over a burned house that a family no longer owns.
The Tool She Thought Was Checking Her Work
Eric Khodadian represents Meni-Siliga. Working through the motions State Farm's defense had filed, he found what he later described in his own filing as cases that do not exist, quotes that do not exist, and holdings that do not exist. He brought the list to a conference at Los Angeles Superior Court on August 7.
The lawyer who had used the AI was Jacquelene Robinson. In a filing she confirmed the count herself: seven case citations across eight filings that, in her words, simply did not exist, along with some incorrect case titles and quotes that could not be found in the cases she had cited.
Her explanation is the part of this story that should worry every managing partner in the country. Robinson wrote that she used an AI tool called Irys. She believed Irys was tied to her firm's subscription to Westlaw, and she believed that connection meant an internal citation check was being run on the output. She was mistaken about both.
There is no rogue model in that sentence. There is no jailbreak, no prompt injection, no exotic failure mode. A lawyer used a research tool, held an incorrect belief about what that tool was wired into, and never confirmed it. The tool did exactly what these tools do when nothing verifies them.
Nobody Above Her Knew Either
Kenneth Katel of Musick, Peeler and Garrett is lead trial counsel on the defense. In his filing he apologized to the court and to the plaintiff, and explained that his co-counsel had told him she used AI to help prepare some of the filings.
"I was not aware that AI had been used, but as lead trial counsel I accept full responsibility for our filings," Katel wrote. He also told the court his firm has updated its policy on AI use. He did not say what the policy now says.
Khodadian's filing described what happened after the August 7 conference in unflattering terms, writing that a State Farm lawyer was visibly enraged, followed him and his co-counsel down the hall, and repeatedly demanded to see the list of mistakes they had found. Katel disputed that characterization, though he wrote that it is accurate to say he was upset given the seriousness of the allegations. Both men declined to answer questions from reporters afterward.
Read the chain again. The associate did not verify the tool. The partner did not know the tool was in the room. The court found out from opposing counsel. That is three supervision layers and none of them was load bearing.
What The Client Said, And What It Would Not Say
State Farm is reviewing what happened. Spokesperson Tom Hartmann said the company expects its external counsel to conform to the highest level of ethical standards and professionalism, including confirming the accuracy of all legal filings.
Two questions went to the company that it did not answer. How many external law firms represent State Farm. And whether the company asks those firms about their AI policies before hiring them.
The silence on the second question is the interesting one. An insurer of that size runs a procurement process for outside counsel with rate cards, billing guidelines, conflict screens and audit rights. Nothing about adding a line item on generative AI use is technically hard. The absence of an answer suggests the line item does not exist yet, at State Farm or at most of its peers, which means the client is relying on the exact supervision chain that just failed.
The Scale Nobody Wants Printed
A database maintained by the researcher Damien Charlotin tracks legal decisions worldwide in which a court found that somebody relied on hallucinated material. As of its August 16 update it holds 1,922 cases. The United States accounts for 1,313 of them. Canada has 211, Australia 98, the United Kingdom 62, Israel 55.
The breakdown by who was responsible is the number that ends the comfortable version of this story. Self-represented litigants account for 1,111 entries, which is the result people reach for when they want this to be a problem of amateurs. Practising lawyers account for 761. Judges account for 28.
By category, 1,601 involve fabricated material, 800 involve misrepresented material, and 519 involve false quotes. Roughly two in five law firms and nearly half of corporate legal departments now report using generative AI, according to a recent survey, and the professional share of that database has been climbing alongside the adoption curve rather than flattening out.
Daniel Ho, the Stanford law professor who has spent years measuring this, co-wrote a paper finding that a couple of companies overstated how completely their legal research products could guarantee citations free of hallucination. Robinson's belief that Westlaw was checking her work behind the scenes did not appear from nowhere. The industry spent two years telling lawyers that the guardrails were built in.
California Already Set The Price
Last year a California appellate court fined Los Angeles attorney Amir Mostafavi $10,000 after finding that 21 of the 23 quotations in his opening brief were fabricated. The Second District certified the opinion for publication, which turns a sanction into a rule other judges can point at. It remains the largest fine a California court has issued over AI fabrication.
Mostafavi told the court he had written the appeal himself and then run it through ChatGPT hoping to improve the draft, without reviewing what came back and without expecting the model to insert citations at all. Robinson's account is a variation on the same theme with a different tool and a mistaken assumption about a subscription in place of a mistaken assumption about a chatbot.
The Part That Does Not Get Fixed By A Policy Memo
Every firm caught in one of these now announces an updated AI policy, and Musick, Peeler and Garrett has done exactly that. Policies are cheap. The failure here was not the absence of a rule against unverified AI output. It was that no step in the workflow required anybody to open the cases and read them before the motion went out the door.
Verification is boring, slow, billable at a rate clients hate paying, and it is the only thing that catches this. A tool that promises to remove that step is selling the removal of the one control that works. Robinson thought she had bought a tool with the check included. What she had bought was a tool that produced text.
The counterargument deserves its due. Citation checking has always been delegated, usually to junior associates and paralegals, and human cite checkers have missed things for a century without anyone calling it a crisis. The difference is volume and confidence. A tired associate misses a bad citation. A model manufactures seven of them, formatted correctly, in a document that reads as though it was written by someone who had done the reading.
Meni-Siliga goes to trial in October. Her house has been gone for a while. The filings arguing against her were partly built out of law that was never written, and the only reason anyone knows that is that her lawyer sat down and checked, one citation at a time, the way somebody at the firm on the other side was supposed to.