
By Jeffrey T. Donner, Esq.
September 23, 2026
The most important danger artificial intelligence poses to legal advocacy is no longer simply that it may invent a case. AI can now generate enormous quantities of superficially sophisticated legal work using real cases, real statutes, polished prose, and recognizable legal terminology. What it cannot supply is the professional judgment necessary to decide whether a particular argument is legally sound, supported by the record, procedurally available, strategically useful, or worth presenting to a court at all. A lawyer can verify every citation in an AI-generated brief and still file a terrible brief.
A September 2026 opinion from Florida’s Fourth District Court of Appeal brings that problem into unusually sharp focus. In Lisandrillo v. Palozzi, No. 4D2026-2262 (Fla. 4th DCA Sept. 16, 2026), the court addressed a series of filings it believed exhibited signs of artificial-intelligence assistance. Unlike the first generation of AI-sanctions cases, however, the central concern was not merely fabricated authority. The court confronted what it described as “AI slop” and “issue churning”: large amounts of legal argument produced without the corresponding exercise of a trained lawyer’s independent professional judgment.
That distinction matters. Generative AI has dramatically reduced the cost of producing legal prose, but it has not reduced the need for lawyers to decide what should actually be said. That problem now appears in several forms at once. Lawyers can rely too heavily on AI-generated drafts. Pro se litigants can flood courts with machine-generated arguments. And clients increasingly arrive at lawyers’ offices carrying lengthy AI-generated briefs, motions, and research memoranda that they expect counsel simply to adopt and file.
The developing lesson is broader than “check your citations.” It concerns the nature of lawyering itself.
What Lisandrillo Actually Involved
Lisandrillo arose from contentious dissolution proceedings in Broward County involving a minor child, reunification therapy, timesharing, a guardian ad litem, contempt proceedings, and related disputes. The Fourth District denied a petition for writ of prohibition seeking review of the trial judge’s refusal to disqualify herself, but the court devoted substantial attention to what it regarded as abusive motion practice and apparent AI-assisted advocacy.
The volume of filings was striking. According to the opinion, counsel filed a 41-page “Urgent Omnibus Application” seeking thirteen forms of relief, followed by an 11-page emergency motion seeking another nine forms of relief. Counsel later filed three motions in limine shortly before a hearing and, after the trial court held the petitioner in contempt, a 33-page motion to disqualify asserting numerous grounds. The resulting appellate proceeding involved a 61-page emergency petition that sought prohibition relief concerning judicial disqualification while also attempting to obtain certiorari review of seven additional rulings.
Length alone, of course, proves very little. Complex litigation sometimes requires lengthy filings, and a long brief is not necessarily a bad brief. What troubled the Fourth District was the relationship between volume and substance. The court identified “indicia of non-human authorship” and described arguments that were convoluted, unsupported by the record, procedurally defective, irrelevant, unpreserved, or frivolous.
The problem was therefore not that counsel lacked arguments. It was that too little professional discrimination appeared to have been exercised among them.
Issue Spotting Is Not Issue Churning
Lawyers are trained to identify legal issues. Experienced lawyers are also trained, through practice if not in law school, to discard most of them.
A competent litigator does not merely ask what arguments can conceivably be constructed from a set of facts. The lawyer must determine which issues actually advance the client’s interests, which have been preserved, which are supported by admissible evidence, which are procedurally available, which survive the governing standard of review, and which are important enough to justify occupying the court’s attention.
Generative AI changes the economics of this process. A user can ask for twenty possible arguments and receive twenty within seconds. The user can then request additional constitutional theories, due-process objections, sanctions grounds, procedural challenges, contrary arguments, and responses to those contrary arguments. The marginal cost of generating more legal text approaches zero.
That capacity is potentially valuable. Lawyers should want tools that help identify issues they might otherwise overlook. But generation is only the beginning of analysis. The professional work begins when someone decides which arguments survive scrutiny.
That is where AI can encourage bad habits. An inexperienced advocate may assume that more arguments make a stronger filing. Experienced litigators usually learn the opposite. Arguments compete for judicial attention. A weak point can diminish the credibility of a strong one. A marginal procedural complaint can obscure a dispositive issue. A brief containing thirty alleged errors may communicate that counsel has failed to determine which errors actually matter.
Much of experienced advocacy therefore consists of subtraction. The motion that would accomplish nothing is never filed. The argument that is technically available but strategically harmful disappears during editing. The issue that consumes ten pages without materially improving the client’s position never reaches the judge.
That invisible editing is not a lack of diligence. It is professional judgment.
The Same Problem Is Arriving From Clients
The reported cases capture only part of the change now occurring in litigation practice. Clients have access to the same generative tools lawyers use, and many have begun using them extensively.
In my own practice, I have repeatedly received AI-generated legal material from clients and prospective clients during roughly the last year and a half. The submissions have included research memoranda, proposed complaints, motions, appellate arguments, constitutional theories, and documents that the client viewed as essentially finished court filings. Some exceeded one hundred pages. I have seen AI-generated submissions approaching or exceeding 170 single-spaced pages.
One prospective client had produced extraordinarily lengthy AI-assisted materials containing scores of proposed arguments and appeared to believe that retaining counsel would mean having a lawyer continue that approach under the lawyer’s signature. I was willing to undertake the representation for an appropriate retainer, but only with the clear understanding that I would decide, as counsel, what arguments were legally appropriate and what documents should be filed. I was not prepared simply to transmit the client’s AI product to the court under my name.
The prospective client did not particularly like that limitation. That reaction illustrates a new source of tension in the attorney-client relationship.
I have also received material formatted as though it were already counsel’s work product, sometimes with my name effectively incorporated into the proposed filing. That presents a separate concern. A document bearing a lawyer’s name can circulate electronically among co-counsel, opposing counsel, mediators, or others and create the false impression that counsel prepared or approved it. A lawyer has a legitimate professional interest in preventing unreviewed client-generated material from being attributed to the lawyer even before it reaches a courthouse.
The essential problem, however, is substantive rather than cosmetic. The fact that AI can make a document look like a brief does not mean that the underlying legal work has been done.
A 170-Page AI Brief Is Not 170 Pages of Free Legal Work
Clients who produce these materials may reasonably believe that they have saved counsel substantial time. The document contains headings, statutes, cases, quotations, and confident legal conclusions. It looks like work that historically might have required many attorney hours.
The lawyer sees something different. Every material authority must be checked. Every important quotation must be read in context. Every factual assertion must be compared with the record. Each theory must be tested against preservation, jurisdiction, procedure, available remedies, contrary authority, and the actual posture of the case.
If a client generates fifty possible theories and forty-five are useless, counsel may still have to spend considerable time determining why those forty-five theories are useless. An AI-generated memorandum can therefore increase the legal bill rather than reduce it because it creates a body of material that would never have required professional review had the case been analyzed in a more disciplined manner from the outset.
This does not make client use of AI inherently undesirable. A client may identify a useful authority, uncover an overlooked fact, or suggest a legitimate question worth investigating. But there is a fundamental difference between providing counsel with useful leads and delivering a completed-looking brief with the expectation that the lawyer will merely sign it.
The latter misunderstands where the lawyer’s value lies.
AI Can Make Weak Arguments Look Strong
One reason this problem is difficult is that modern generative systems are extraordinarily fluent. Before AI, a nonlawyer attempting to produce a sophisticated legal brief often revealed fairly quickly that the author did not understand procedure, precedent, or the structure of legal analysis. Modern AI can conceal those deficiencies behind polished prose.
A generated document may cite actual cases, quote constitutional provisions, use sophisticated terminology, and describe legal standards with great confidence. To a nonlawyer, it may look authoritative. Yet the legal theory may depend on an unpreserved issue, an unavailable remedy, an inaccurate characterization of the record, or an authority arising in an entirely different procedural setting.
A client may consequently arrive believing that AI has identified thirty devastating errors while counsel concludes that three issues deserve serious attention. The lawyer can then appear insufficiently aggressive simply because the machine generated more.
But producing a theory is not the same thing as evaluating one. Trial and appellate experience teach lawyers that credibility matters, procedural posture matters, the audience matters, and selecting the strongest issue frequently matters more than identifying the greatest number of issues.
AI can generate the vocabulary of advocacy. It cannot assume responsibility for exercising that judgment.
Sometimes the Lawyer Must Tell the Client No
The lawyer’s obligation to exercise independent judgment means that counsel cannot simply function as the client’s electronic filing service.
A client controls important objectives of the representation, but the lawyer remains professionally responsible for the means by which the representation is conducted. If a client hands counsel a sixty-page AI-generated motion and demands that it be filed, the lawyer still must determine whether filing it is legally and strategically appropriate. If an argument is frivolous, attributing it to the client does not make it permissible. If a factual allegation lacks record support, the fact that the client’s AI system generated it does not cure the defect.
At some point, professional independence may require counsel to tell the client that an argument will not be made, a motion will not be filed, or a proposed document will not bear the lawyer’s name. If the disagreement is fundamental and cannot be resolved, declining or ending the representation when ethically and procedurally permissible may be preferable to surrendering professional judgment.
That principle is not new. AI simply makes the conflict more common because clients can now arrive with something that looks remarkably like completed legal work.
The Pro Se Cases Show What Happens Without a Gatekeeper
Florida appellate courts have already confronted similar problems when no lawyer stands between the AI output and the court.
In Avery v. Beauzil, the Fourth District addressed a pro se litigant who submitted a lengthy, single-spaced AI-generated extraordinary-writ petition and engaged in extensive motion practice. When ordered to explain why filing restrictions should not be imposed, the litigant responded with still more extensive AI-generated material. The court ultimately restricted further pro se filings unless signed by a Florida lawyer.
In Gouldy v. Chiasson, another pro se appellant raised numerous issues and subissues. The Fourth District identified a nonexistent case, fabricated quotations attributed to actual decisions, misrepresentations of the record, and authorities that did not support the propositions for which they were cited. The court also observed that any potentially legitimate argument was overwhelmed by the quantity of meritless ones.
I have seen the same dynamic outside the reported decisions. In my practice, I encountered a prospective client who sought professional assistance only after filing an extraordinary, approximately 250-page pro se complaint and facing judicial scrutiny over what she had done. The complaint itself disparaged the need for lawyers and suggested, in substance, that modern litigants could dispense with them. Only after the consequences of self-directed litigation became apparent did the litigant seek counsel.
The larger point is structural. AI can give a nonlawyer the ability to produce the volume and outward appearance of legal work without giving the person the corresponding ability to determine whether the material should ever be filed.
When the material instead reaches a lawyer before filing, counsel becomes the final professional filter between machine-generated advocacy and the judicial system.
The Record Presents a Separate AI Risk
Lisandrillo also illustrates a danger beyond legal research. The Fourth District found that several allegations in the filings were inconsistent with the hearing transcript and expressed concern about arguments apparently derived from errors contained in an unofficial rough transcript.
That issue deserves particular attention because litigation records are inherently imperfect. Witnesses contradict themselves. Transcripts contain mistakes. Documents are incomplete. Chronologies are disputed. Generative AI, meanwhile, is designed to transform incomplete information into coherent language.
That ability can cause uncertainty to disappear. An inference becomes a factual assertion. An ambiguous exchange becomes a definitive ruling. A transcription error becomes evidence of judicial misconduct.
Lawyers therefore must distinguish between an AI-generated description of the record and the record itself. AI may help organize, summarize, and analyze transcripts and exhibits, but the transcript remains the transcript and the exhibit remains the exhibit. Material factual assertions must ultimately be verified against the underlying source.
The Economics Affect Courts as Well as Clients
AI has dramatically reduced the cost of generating legal prose. It has not similarly reduced the cost of reviewing it.
A party can generate fifty pages overnight. Opposing counsel must still analyze them. A litigant can produce ten additional theories within seconds. A judge still must determine whether any one of them has merit. The production cost approaches zero while the evaluation cost remains substantial.
That asymmetry is one reason issue churning is not merely a problem of bad writing. It can impose real costs on opposing parties and courts.
The same principle explains the billing problem when client-generated AI reaches counsel. Raw text has become inexpensive. Professional evaluation has not. A lawyer should not bill hypothetical hours that technology eliminated, but neither should anyone assume that an instantaneous AI draft requires no professional labor afterward. The responsible lawyer still must verify, edit, research, eliminate, and ultimately decide what can ethically be presented under counsel’s name.
AI makes words cheap. It does not make responsibility cheap.
The Lawyer’s Signature Remains the Boundary
Florida’s existing professional rules already provide much of the necessary framework.
Florida Rule of General Practice and Judicial Administration 2.515 treats a lawyer’s signature as a professional representation concerning the filing. In 2026, the Florida Supreme Court amended the rule to make explicit that legal authorities identified in a filed document exist and are accurately cited. The Rules Regulating The Florida Bar separately impose duties concerning competence, independent professional judgment, meritorious contentions, candor, expedition of litigation, and reasonable fees.
Those obligations do not depend on who produced the first draft.
If the lawyer personally typed every word, a frivolous argument remains frivolous. If AI drafted the document, a sound argument remains sound after appropriate professional review. If the client generated the material and demanded that counsel file it, the lawyer nevertheless owns the filing once the lawyer signs it.
Courts therefore need not solve every question of AI authorship. The more important question is whether the lawyer fulfilled the professional obligations that already attach to advocacy.
The Procedural Posture in Lisandrillo Matters
The seriousness of Lisandrillo should not be overstated beyond what the Fourth District actually did.
The court denied the requested extraordinary relief and ordered petitioner’s counsel to show cause within ten days—expressly without the use of AI—why sanctions should not be imposed. The court identified possible referral to The Florida Bar for disciplinary consideration among the potential consequences.
As of the September 16, 2026 opinion, however, that was an order to show cause, not a final disciplinary adjudication. The distinction matters, particularly when discussing another lawyer’s conduct publicly.
The broader significance of the opinion does not depend on what ultimately occurs in the sanctions proceeding. The Fourth District already made clear that legal advocacy can become professionally problematic even where the authorities cited are real. The quality of lawyering cannot be measured merely by whether the cases exist.
The Scarce Resource Is Judgment
The first generation of AI ethics cases taught lawyers a straightforward lesson: verify the authorities.
The next lesson is more important. Legal prose is no longer scarce. Potential arguments are no longer scarce. A generative system can produce motions, objections, constitutional theories, appellate issues, sanctions arguments, and counterarguments almost without limit.
Judgment remains scarce.
The value of an experienced litigator therefore does not disappear as AI becomes better at producing legal work. In some respects, it becomes easier to identify. When anyone can generate fifty arguments, the professional value may lie in recognizing that forty-seven should never reach the judge.
That judgment includes knowing when the record does not support the client’s description of events, when a theoretically available argument distracts from a stronger one, when aggressive rhetoric damages credibility, when a purported emergency is not an emergency, and when a client’s impressive-looking AI brief is something no responsible lawyer should sign.
Generative artificial intelligence can make good lawyers more efficient and potentially more effective. It can help identify issues, organize records, test arguments, compare authorities, and expose weaknesses that might otherwise be missed. Used correctly, it should expand the range of possibilities a lawyer can consider before narrowing them to the best professional product.
But AI does not eliminate the lawyer’s central responsibility. It makes that responsibility more important.
The question is no longer simply whether the cases are real. The question is whether the lawyering is.

Some months ago I wrote a short AI program/instruction set that would take an input document (think motions) and verify all the cited cases. The problem AI and I agreed on was assigning the AI quality control process to include logic validation. That was simply too challenging.
Your article (AI Slop), gives a new impetus to the theme of AI analysis of legal reasoning. That is a distinct category: an unsupported proposition rather than a fake authority. Such errors (corrupted input) can be built upon such that legally fluent nonsense is produced. That is, AI can create fluff, but can’t be trusted to find it. Sorting wheat from chaff, identifying dispositive issues, etc. is still the job of the human.