On September 16, Florida’s Fourth District Court of Appeal denied a writ petition in a divorce case, Lisandrillo v. Palozzi, No. 4D2026-2262, and used its opinion (opens in new tab) to write about something else: what it called “AI slop.” It ordered the petitioner’s lawyer to show cause within ten days, “without the use of AI,” why sanctions such as a referral to The Florida Bar should not be imposed for frivolous filings.
Most AI sanction stories are about fabricated citations. This one is about reasoning, which is why I would put it in front of any firm whose AI policy amounts to “check every cite.”
What the Court Found
The court’s first footnote says the problem “has nothing to do with hallucinated law or false legal citations,” and it quotes a Bloomberg Law commentary by Jodi Swick and Jarrett Ganer: “The citations are real. The cases exist. The language is confident.” The court flagged one familiar AI error, the claim that a motion to disqualify automatically stays all proceedings in the trial court, but it identified no invented cases. What it described instead was a pattern:
- A 61-page emergency petition asking the court to review eight separate rulings at once, which the court declined to treat as an emergency.
- A 33-page motion to disqualify the trial judge on 14 grounds, many of which, the court said, were contrary to the transcript.
- Accusations drawn from an unofficial “rough draft transcript” that the official transcript did not support. The court wrote that it suspected “AI generated these bogus arguments from mistranscriptions.”
- What the court called “issue churning”: “AI churns out a laundry list of convoluted arguments that no competent lawyer should make.”
The court said it suspected AI was used; it made no finding, and a footnote observes that some passages “appear too unpolished to have been produced by AI.” And the lawyer told Law.com, as the ABA Journal reported (opens in new tab), that she does not use AI to generate court filings. She will be heard in her show-cause response. My interest here is the rule the court stated.
The Sentence That Matters
The court put the rule in two sentences: “[W]hether this was AI-generated, AI-assisted, or something else makes no difference. An attorney is obligated to stand behind any submission to a court.”
The court anchored that in Florida’s signature rule. By signing, counsel represents that she has read the document, that “there are good grounds to support” it, and that it “is not interposed for delay.” The court then listed the professional conduct rules the filings implicated: competence, the adviser’s duty of independent professional judgment, meritorious claims, expediting litigation, and candor to the tribunal. None of those rules mentions AI, and the court did not need one that did.
The opinion also raises billing, which I have rarely seen in these cases. The first factor in a reasonable fee is the time, labor, and skill the work required, and the court observed that “[i]t takes almost no skill and little time to generate AI slop.” A firm that bills drafting hours for text no lawyer really drafted has a second problem on its hands.
Why a Cite-Check Would Have Missed This
A citation check would have passed almost everything in these filings. What failed was upstream of the citations. The arguments came from a source outside the official record, and no one appears to have decided which of them were worth making.
When I work with firms, I describe four controls for AI-assisted legal work: a closed set of sources, accuracy measured before anything is trusted, verification that leaves a record, and a named lawyer who signs off. This opinion is about the first and the last.
A closed set of sources. A tool that helps draft a filing should work from a set you control: the record, the file, the governing authority. An unofficial rough transcript is a lead to check against the official one. If a factual claim traces back only to the rough draft, the review should catch it before the court does.
A named lawyer who signs off. The signing lawyer has to decide what the argument is. Fourteen grounds for disqualification is what a filing looks like when that decision never gets made. The review that matters asks whether each argument deserves to be in the brief, which takes longer than confirming that each case exists.
The middle two controls apply as well. Measuring a tool’s accuracy can and should include its reasoning. When I used a language model to code trademark decisions for my research, I checked 1,002 of its factor-level findings, one at a time, against my own reading of the decisions. I was the only reviewer, and I could see the model’s answers, so I report that limit with the result. A check like that tests whether the model understood the decision, which is the question a citation check never reaches. And a review that leaves a record lets a firm show, when a court asks, who looked at the filing and what they checked.
What I Would Take From It
For firms and legal departments:
- Give AI tools the official record. Treat anything else, including rough transcripts and client summaries, as material to verify first.
- Make the pre-filing review about the argument. The signing lawyer should be able to say, for each ground, why it is worth raising.
- Keep a record of who reviewed what, so the certification on the signature line has something behind it.
- Look at time entries for AI-assisted drafting. The Florida court tied the quality of the work to the reasonableness of the fee.
The Fourth District said that courts “must require more from attorneys and not tolerate abusive practice fueled by AI.” The opinion is not final until any timely rehearing motion is resolved, and the show-cause response is still to come. I will update this post when the court acts on it.
Prof. Reichert