Yesterday, in AI Found an Insurance Coverage Issue the Researchers Missed: What That Means for Legal Scholarship, I wrote about how artificial intelligence identified an insurance coverage issue that experienced human researchers had missed. Several AI reasoning models offered a perspective on policy language that researchers had not fully appreciated. The researchers then did exactly what professionals should do: they went back to the policy, checked the analysis, reconsidered their conclusion, and corrected their scholarship. AI served as a powerful second reader.

A few hours later, the Los Angeles Times reported the other side of the same technological revolution. 1 Attorneys representing State Farm in wildfire cases acknowledged that court filings in a California insurance dispute contained nonexistent case citations apparently generated through artificial intelligence. Attorney Jacquelene Robinson of Musick, Peeler & Garrett stated that she had used Irys, a legal AI product, and failed to verify the citations before the material was filed. She accepted responsibility for what she called a serious lapse in professional judgment.

Both stories demonstrate that the use of AI in the legal profession requires more understanding than simply saying, “AI hallucinates.” In the first, humans used AI to challenge their thinking and then checked its work. The result was better scholarship. In the second, AI-generated legal material reached a court without adequate verification. The result was a worse legal product and a professional embarrassment.

The important difference was not the algorithms and systems. It was the professional process in the use of those systems.

I learned something about citation checking long before artificial intelligence entered the legal profession. When I was the Executive Editor of the Law Review at the University of Florida, I spent hundreds of hours checking citations in articles submitted by professors from law schools around the country. There were mistakes. Some professors had a lot of them.

Citations did not always support the propositions for which they were offered. Quotations were sometimes inaccurate. Words were omitted. Context mattered. Occasionally, after reading the actual authority, I would wonder how the cited material had been characterized the way it was.

What almost never happened was somebody simply inventing a case.

Inaccurate legal citation did not begin with AI. Lawyers, professors, judges, law clerks, and law students have always made mistakes. Human fallibility is hardly new.

What artificial intelligence changes is the speed and scale of those mistakes. A lawyer could always write a bad brief or mischaracterize language from a case. AI can help that lawyer create a much longer bad brief in a fraction of the time, beautifully organized and confidently written, with legal authority that appears perfectly legitimate.

That may be the greater danger. We spend a lot of time discussing how much time AI saves. But saving time is not necessarily the same as improving legal work.

Suppose research and drafting that once took five hours can now be accomplished with AI assistance in two. What happens to those three saved hours? A good lawyer can use that time to read the important cases, search for contrary authority, study the factual record, test the argument, examine policy language, anticipate opposing counsel, simplify the writing, and question whether an initial conclusion is correct.

Another lawyer can simply produce three more briefs. The first lawyer has used technology to improve professional work. The second may simply be making more junk faster.

There is another reason I think the focus on hallucinations or fake citations misses a significant danger. A nonexistent case is one of the easier AI errors to discover. Search for it. If it does not exist, you have your answer.

The more dangerous error is the case that does exist. The citation is real. The case may even be good law. But it does not stand for the proposition the AI offers. Perhaps the facts are materially different. Perhaps the language is dicta. Perhaps another paragraph limits the holding. Perhaps the jurisdiction or procedural posture makes the case irrelevant. Perhaps AI has taken a quotation out of context.

A machine citation checker cannot practice law. That takes professional judgment.

Ironically, Irys’s own documentation recognizes this distinction. Its Cite Check feature can verify whether a citation exists and remains good law, but the company warns that this does not mean the authority actually supports the proposition for which it is cited. The lawyer still must read the case.

We do not yet know whether Irys’s citation-checking feature was used in the State Farm matter or precisely how the errors escaped review. Those facts matter, and it would be unfair to blame the software company without knowing them. What we do know is that any lawyer who uses such technology must verify the accuracy of its results.

I have been licensed in California for about 15 years. California’s ethical guidance on artificial intelligence is clear that lawyers must understand the capabilities and limitations of the systems they use and retain control over professional judgment. The Ninth Circuit recently said much the same thing while sanctioning lawyers in another AI-related citation matter. The ethical problem is not that an attorney used AI in research or drafting; the problem is when an attorney signs and files a document without ensuring all AI generated work is accurate.

None of this changes my view that good lawyers should use AI aggressively. Ask it what you missed. Ask it to attack your argument. Ask it to find cases that counter your argument and undermine your position. Ask whether another policy provision changes the result. Ask it to organize enormous amounts of material and identify inconsistencies a human might overlook. I remain excited about this technology and know it helps our clients.

But AI’s answer should begin the inquiry, not end it. If a case matters, open it and read it. If a quotation matters, find it in the opinion. If policy language matters, read the entire provision along with the definitions, exclusions, exceptions, conditions, and endorsements that affect it.

If AI says there is no contrary authority, do not simply ask the same machine, “Are you sure?” and call that verification. That is the technological equivalent of asking a witness, “Are you sure?” and ending the cross-examination when the witness says yes.

Verification means returning to the source and exercising independent professional judgment.

Artificial intelligence should not cause lawyers to think less. Its promise is that machines can increasingly handle mechanical work, so we have more time to do what clients hire lawyers to do: think, question, judge, strategize, communicate, and get things right.

Use the machine aggressively. Trust it cautiously. Verify it relentlessly.

Keep the judgment human. Otherwise, this remarkable technology may simply allow us to make more mistakes faster.

Thought For The Afternoon

“Every technology is both a burden and a blessing; not either-or, but this-and-that.”
—Neil Postman, Technopoly: The Surrender of Culture to Technology (1992)

1 Rebecca Ellis, “AI hallucinated case law in insurance company’s filings in L.A. County house fire dispute,” Los Angeles Times, Aug. 19, 2026. Available online at https://www.latimes.com/california/story/2026-08-19/ai-hallucinations-case-law-state-farm-la-county-fire-dispute (may require subscription).