AI for Lawyers: Checking Citations Is a Duty That Does Not Turn on Intent

TL;DR: Two July 2026 decisions make the same point from opposite ends. The Arizona Court of Appeals sanctioned a self-represented filer for AI-invented citations and held that not meaning to deceive is no defense. A federal judge in Michigan warned government lawyers on the record over a citation she could not find. The duty to check AI citations does not depend on intent or on who you are. The only protection is a checking step that runs before anything is filed.

The most useful lesson in AI for lawyers this year is not about which tool to buy. It is about who owns the check. Two July decisions, one from a state appellate court and one from a federal trial court, landed on the same answer: the person who files owns it, whether they meant to mislead or not. I run a multi-agent AI team applied to legal work, business projects, and teaching others how to use AI, and the rule my agents work under is the one these courts applied.

What did the Arizona Court of Appeals decide in Dineen/Shibata v. Kotchka?

On July 15, 2026, the Arizona Court of Appeals, Division One, issued a published opinion in Dineen/Shibata v. Kotchka, No. 1 CA-CV 25-0606 PB. Judge Brian Y. Furuya wrote it, joined by Presiding Judge Andrew M. Jacobs and Judge James B. Morse Jr. A self-represented appellant in a probate appeal filed a brief built with AI research tools. The Arizona Mirror reported that six of the eight citations were deficient and two of the cases did not exist. The court held that citing a hallucinated case is sanctionable, and that a lack of intent to deceive is no defense to failing to check.

The court also said no filer can hand off the duty to make sure their citations are accurate to a technology tool, AI included. The appellant had no legal training and did not mean to mislead anyone. Neither point excused him. They only shaped which sanction fit. The sanction was the part of the other side's attorney's fees caused by the fake citations, plus costs.

Does intent matter when AI invents a case?

Not to whether there is a violation. Under the Arizona opinion, intent goes to the sanction, not to the duty. That makes sense once you look at what the duty is. It is not a duty to be honest about your sources. It is a duty to check them. Checking does not depend on your state of mind. Either you pulled the case and read it, or you did not. A filer who believed every citation was real and never checked has still failed the duty, because belief was never the standard.

This is the part that gets lost in the "ChatGPT lawyer" coverage. The stories focus on the embarrassment of a fake case. The legal point is narrower and more useful. Good faith is not a substitute for a checking step.

What happened in Daghra v. Hinkley?

The second decision came a day later, from the other end of the spectrum. On July 16, 2026, in Daghra v. Hinkley, No. 1:26-cv-01429, an immigration habeas case in the Western District of Michigan, Chief Judge Hala Y. Jarbou issued an order addressing the government's brief. The brief cited "Taylor v. Hott, 724 F. App'x 387 (6th Cir. 2018)." The court could not find that case or the quoted language anywhere. According to Newsweek's report of the order, Chief Judge Jarbou wrote that the citation was likely produced by generative AI, said she would not presently impose sanctions, and warned that AI research has fueled "a rash of cases" with fabricated authorities.

The order dismissed the petition without prejudice and included a notice about the fictitious citation. The government lawyers were warned on the record. They were not sanctioned.

Why does it matter that one filer was self-represented and the other was the federal government?

Because it shows the duty does not depend on who you are. At one end, a person with no legal training, working alone, was sanctioned. At the other end, federal government lawyers, with every institutional resource behind them, put a citation no one could find in front of a federal court and got a written warning in a public order. Training did not prevent the error. Lack of training did not excuse it. In both cases the failure was the same: nobody pulled the citation and read it before the brief went in.

That is the uncomfortable point for anyone thinking about AI for law firms of any size. Resources do not check citations. People and processes do.

How many judges are using AI themselves?

More than you might expect, at least among those who answered. In a survey Northwestern released with the New York City Bar Association on March 30, 2026, 112 of 502 sampled federal judges responded. About 60 percent of the judges who responded had used at least one AI tool, and about 22 percent used one weekly or daily. The response rate means the numbers describe the judges who answered, not the federal bench as a whole.

Still, it changes the picture of who is reading these briefs. Many of the judges who responded know what these tools do and what they get wrong.

What does Florida's rule require?

Florida wrote the duty into its rules this year. The Florida Supreme Court adopted an amendment to Rule 2.515(d)(2) on May 28, 2026 (SC2026-0673), effective June 15, 2026. As adopted, each signer of a filing represents that the legal authorities cited exist and are accurately cited. It applies to lawyers and self-represented filers alike, and courts may impose sanctions after notice and an opportunity to be heard. The rule does not require anyone to disclose AI use. A companion administrative order, AOSC26-12, bars local AI disclosure or certification requirements, so the statewide rule replaces the patchwork.

The Florida rule and the Arizona opinion arrive at the same place by different routes. The signer represents that the citations are real. How the brief was drafted is beside the point.

What does a checking step look like in a working AI system?

It is a separate step, owned by a person, that runs before anything goes out. In the system I run, drafting and checking are different jobs. An agent can draft. A citation stays unverified until the source has been pulled from a primary source and the cited passage has been read against what the draft says it stands for. Anything that fails goes back flagged, not quietly fixed. A person reads the result before it leaves the system. The tool does not own the check. The person signing off does.

None of that is complicated. The hard part is making it non-optional on the days when there is no time. I laid out the structure in more detail in a framework for verifying AI-generated legal content. For how the sanctions picture built up from the first widely reported case to this year's rules, see the post on Mata and the regulatory regime that followed. And for the habit that comes first, why I stopped asking ChatGPT for legal citations at all.

What should anyone filing with AI help take from this?

Three things.

One. Intent will not save a bad citation. Courts are treating the failure to check as the violation, and good faith as relevant only to the sanction.

Two. Who you are will not save it either. The duty runs to the self-represented filer and to the federal government alike.

Three. Build the check into the process, not into good intentions. AI legal research tools will keep improving, and they will keep getting some citations wrong. Whatever you use, whether it is ChatGPT, Claude for lawyers, or a research product built for the purpose, the step that matters is the one where a person reads the source.

No court has accepted that an artificial intelligence lawyer drafted the brief and so nobody is responsible. Every one of them has looked for the person who signed it.

Until next week, the team keeps running.