The Citations Were Real. That Was Not Enough.

The 82-page response missed the court's point

The case, Lisandrillo v. Palozzi, arose from a dissolution proceeding involving a ten-year-old child, reunification therapy, visitation, a guardian ad litem, and contempt litigation. The appellate court denied an emergency petition for extraordinary relief on September 16. It then ordered petitioner's counsel, Jaclyn Soroka, to explain why she should not be sanctioned for what the court called an abusive filing. The panel suspected that the filings contained AI-generated components, but it repeatedly said the source of the language was not the decisive question.

Soroka disputed the court's characterization. Reuters reported that she denied submitting frivolous material and said she used AI as a modern practice tool but had not used an AI program to generate a court filing. In her response, she emphasized that her cases were real, her citations were valid, and she had personally adopted the words. The October 21 panel said the response did not deny AI use, evaded the issue, and attempted to defend arguments the prior opinion had deemed frivolous.

The new order did not impose a fine or adjudicate a disciplinary violation. It referred the matter to The Florida Bar, which has its own process and standards. The court also noted that its order was not final until disposition of any timely motion for rehearing. Those procedural limits matter. A judicial referral is consequential, but it is not a bar judgment and should not be reported as one.

What the panel did decide is still significant. It rejected the idea that personal adoption completes the lawyer's duty. Responsibility for a signed filing includes responsibility for what the document does, not simply proof that a human looked at its sentences. The court's answer to an 82-page defense was two pages. The difference in length is part of the lesson.

A lawyer's signature cannot convert an unjudged argument into independent professional judgment.

The alleged failure was selection, not just accuracy

The September opinion provides the detail missing from the short referral order. The court described a 41-page urgent omnibus application seeking thirteen forms of relief, an eleven-page emergency motion seeking nine forms of relief, a 33-page judicial-disqualification motion asserting numerous grounds, and a 61-page emergency petition seeking review of eight separate rulings. Volume was not the violation by itself. The court's concern was that the documents bundled incompatible procedures, repeated arguments, elevated nonemergencies, and obscured the relief the tribunal could lawfully provide.

The panel identified assertions it said were contrary to the hearing transcript. It said the petition claimed counsel had been prevented from moving to disqualify when the record showed no such attempt. It rejected an allegation that the judge had continued while counsel's audio failed, noting that the transcript instead showed the judge alerting counsel that she was muted. It criticized reliance on an unofficial rough transcript to claim that the judge had mocked counsel by mispronouncing her name, when the rough transcript misspelled several participants' names many times.

The October order added a different problem. Counsel acknowledged that a lower-court motion contained a lengthy false quotation attributed to a Fourth District case. She described it as human error caused by drafting under time pressure. The panel treated that explanation skeptically. Still, the most important feature of the case is that the court did not reduce its analysis to a fake quotation. It focused on the larger act of advocacy: identifying the dispositive issue, grounding it in the official record, using the correct vehicle, and omitting claims that do not advance the client's lawful objective.

That is why accurate citations are necessary but insufficient. A case can exist and be irrelevant. A quotation can be exact and answer the wrong question. A legal proposition can be true in the abstract and unavailable because it was not preserved. A record citation can point to words that do not establish the event counsel says occurred. Verification confirms components. Judgment decides whether the components belong together.

Florida's filing rule is a floor, not the whole building

Florida sharpened its filing rule this year. Effective June 15, amended Rule 2.515 provides that a signer represents that the document has been read, that good grounds support it, that it is not interposed for delay, and that the identified authorities exist and are accurately cited. The Florida Supreme Court expressly authorized sanctions after notice and an opportunity to be heard for filings inconsistent with those representations.

The citation provision responded to hallucinated authority, but the other representations remain equally important. Reading is not the same as analyzing. Existing authority is not the same as good grounds. Accuracy does not establish purpose. A filing can pass the citation test while failing because the argument lacks factual support, misuses procedure, multiplies issues, or appears designed to delay.

Florida Bar Ethics Opinion 24-1, which is advisory rather than binding, reaches the same point from professional responsibility. It permits lawyers to use generative AI but says they remain responsible for work product and professional judgment. It connects AI oversight to competence, meritorious claims, candor, truthfulness, supervision, and the nondelegable functions requiring personal judgment and participation.

The Bar's Professionalism Expectations make the subtraction problem explicit. Lawyers should pursue client objectives economically and expeditiously, avoid frivolous arguments and delay, withdraw claims that are superfluous or cumulative, and exercise independent judgment even when a client wants a more aggressive course. Some expectations restate mandatory rules; others are recommendations. Together they describe advocacy as disciplined choice, not maximal output.

The real review protocol begins with deletion

Many law-firm AI policies are built around a citation audit: confirm the case exists, check that the quoted language appears, and update the authority. Those are sound controls. Lisandrillo shows that they test only whether the machine supplied plausible building materials. They do not test the architecture.

A substantive review should begin with the requested relief. What can this court do through this procedural vehicle? Then test each proposed argument against the official record, preservation history, governing standard, and remedy. If the argument is true but does not change the result, it probably does not belong. If it depends on an inference from a rough transcript, verify it against the official record. If it attacks seven orders in a proceeding limited to one, separate the issues or abandon the unavailable ones. If an emergency label is necessary only to move ahead of other litigants, the label is a warning, not a strategy.

The final pass is not another search for missing content. It is a deletion pass by a lawyer with authority to say no. Remove cumulative grounds. Collapse overlapping sections. Cut adjectives that substitute for proof. Eliminate record characterizations that cannot be tied to a page and line. Ask whether the strongest argument becomes harder to find because weaker arguments surround it. AI can generate alternatives cheaply; it cannot decide how much institutional attention the client is entitled to consume.

This discipline also protects clients from the economics of abundance. More pages require more review, invite more opposition, raise fees, and create additional opportunities for inconsistency. In a family matter involving a child and contested visitation, procedural sprawl does not remain on the lawyer's screen. It can extend conflict and divert attention from the decisions the court must actually make.

AI makes additional arguments nearly free to generate. It does not make them free for the client, the opponent, or the court to absorb.

The profession has been measuring the easiest defect

Hallucinated citations are measurable. They support a clean rule and a simple headline. Either the authority exists or it does not. The harder failures occupy the space where lawyers earn their role: deciding which facts matter, which procedures fit, which concessions improve credibility, and which arguments should never leave the draft.

The Florida referral does not prove that generative AI produced the filings, establish a bar violation, or settle the underlying family dispute. It does show a court refusing to let the AI question collapse into citation hygiene. Counsel's stated human review did not end the inquiry because the court was evaluating the quality and consequences of the judgment reflected in the filing.

That is the standard law firms should adopt before another tribunal writes it for them. Verify every authority. Then perform the more difficult review: Is it accurate to the record? Is it procedurally available? Is it material to the remedy? Does it sharpen the client's position? And, most importantly, should it be filed at all?