Executive Summary

Key Figures

The four numbers come from four different cases. The first two point to where the force that pulls chatbot records into court comes from, and the last two set the one place that force has been stopped beside the place it has never been stopped at all.

Sources: Volokh Conspiracy (October 19, 2026), eDiscovery Today (October 12, 2026), BSF Perspectives.

It began in the lobby of the Fontainebleau in Miami Beach in December 2024. A guest and Rabbi Shmuley Boteach fell into an argument about the war in Gaza. Boteach, who calls himself America's Rabbi, has 1.2 million social media followers. Over the following weeks he publicly accused the hotel and its staff of tolerating antisemitism, and the company that owns the hotel sued him for defamation and tortious interference in the Southern District of Florida. The case number is 1:25-cv-20251.

The case went badly for the defendant. In September 2025 the district judge denied his motion to dismiss, so the suit went forward, and in February 2026 the magistrate judge imposed sanctions. His answer and defenses were struck, his counterclaims were dismissed with prejudice, and liability was entered for the plaintiff. That leaves a single question for the trial set for February 2027. How much the hotel collects.

It is at that stage that the plaintiff asked for the defendant's AI chatbot records. Its filing describes the prompts and conversations as a written distillation of the defendant's state of mind — no better window exists into what a person actually meant, the argument runs. The headline on the October 19 Volokh Conspiracy post that reported the demand took the plaintiff's framing straight from the brief: the AI data will provide an unrivaled view into the mind of a defendant.

The defense answered that the request sweeps far too broadly. U.S. civil procedure has a discovery phase in which each side must hand over the material it holds before trial, and that phase runs under a deadline. The general deadline had already passed, the defense says, and the plaintiff is now using the damages stage to ask for everything exchanged with an AI. Which side is right has not been settled. The court has neither granted nor denied the request, and has not said how far it would reach if granted. The one fact fixed at this point is that the demand was made.

Why the plaintiff is trying to prove a state of mind at all becomes clear from the structure of American damages. There are two tracks. Compensatory damages make good the loss actually suffered. Separately, punitive damages punish the defendant and deter a repeat. The second track can run high, so states place caps on it. Florida does too.

One exception sits inside that cap. Florida Statutes § 768.73(1)(c) does not apply the cap where the defendant is found to have acted with a specific intent to harm. Prove that he was not merely reckless but set on hurting the other side, and the ceiling on the number disappears. That is why the plaintiff is digging into what was in the defendant's head at the time of the posts. In this case his intent is not an ethical question but a financial one.

The plaintiff rests on more than the state statute. It says it also relies on Supreme Court and Eleventh Circuit precedent weighing a defendant's mental state in punitive damages. It does not say which cases, and the report of the demand does not name them either.

This is where chatbot records occupy a place of their own. A social media post is written for an audience, so it comes dressed up. What goes to a chatbot is written to get a result. You have to set down what you want to do and how it might be done, or nothing useful comes back. The gap between the public statement and the private question is exactly what the plaintiff is after.

One decision from the same autumn reads as the exact opposite. In a case against the Las Vegas police department (Villanueva v. Las Vegas Metropolitan Police Department, 2:24-cv-00125), it emerged that Tom Melton, an expert witness for the defense, had used ChatGPT while writing his report and had not disclosed it. The magistrate judge ordered him to produce the entire ChatGPT log, running to roughly 40 pages.

On September 29 Judge Anne Traum reversed that order, holding that the part requiring everything was clearly erroneous or contrary to law. She split the log in two instead. Facts and data the expert consulted in forming his opinion, such as AI-generated case summaries, are disclosable. The other side has to be able to test that basis, and in particular to check whether fabricated citations found their way in. The expert's own questions and prompts, by contrast, she treated as protected.

What divided the two halves was what the questions reveal. What an expert asked first and what he kept going back to carries how he reads the case and where he senses weakness. The court treated that as mental impressions and theories about the litigation, and extended the purpose of Federal Rule of Civil Procedure 26(b)(4)(B), which shields draft expert reports, to AI conversations. Material prepared for witness examination is protected on the same logic.

The reach of this decision should be read narrowly. It concerns the working process of an expert retained for litigation, not a holding that everything a party writes to a chatbot is protected. It also matters that this was a district judge correcting a magistrate judge's order. It records a disagreement inside a single court about how much to make a party turn over.

One court is reaching for chatbot conversations and the other is keeping them back, which looks like a contradiction. Compare the questions the two cases ask and they do not overlap. In the Fontainebleau lawsuit the issue is what the defendant had in mind when he posted. Intent is itself an element, so the words he wrote at the time are evidence as they stand. In Villanueva the issue is the conduct of the police, not how the expert reasoned. Prying into the path an expert took to his conclusion is closer to looking into the other side's trial preparation, and that territory was protected long before AI arrived.

Reduce the line courts have drawn so far to a single sentence and it comes out like this. Whose mind is this record being used to prove? If it is the party's own mind, it falls on the evidence side; if it is the working process of the side preparing the case, it falls on the protected side. What decides it is not what was asked but whose thinking the question exposes.

Earlier cases laid this line down piece by piece. Over the course of this year U.S. courts have issued one decision after another on chatbot records, and they have tended in one direction. In January the Southern District of New York ordered 20 million de-identified ChatGPT conversations produced in a copyright case, reasoning that records a user handed to a platform voluntarily are not the same as communications intercepted in secret. In February the same court, in Heppner, declined to extend attorney-client privilege on the ground that an AI is not a lawyer. The fact that the terms of service collect the conversations, it held, undercuts any expectation of confidentiality.

March brought a case in which such records landed in the text of a judgment. Fortis Advisors v. Krafton, in the Delaware Court of Chancery. When Krafton acquired the studio behind Subnautica in 2021 it promised 500 million dollars up front and up to 250 million more in earnout, and the chief executive's conversation with ChatGPT about how to avoid paying it was quoted verbatim in the opinion. The exchanges continued after his advisers warned that the legal ground was thin, and the record even holds his complaint when the chatbot replied that canceling the contract would be difficult. He admitted deleting some of the logs, but copies shared over the company's internal messenger survived.

These are decisions by different courts on different questions, and yet lining them up by date shows where chatbot records moved in the space of a year. In January the conversations piled up on a platform were called in wholesale, and by March they were being set into the body of an opinion.

Protection is not entirely absent. In May 2025 the Northern District of California, in the suit between Concord Music Group and Anthropic, left room to treat undisclosed prompts entered at counsel's direction as work product. It did hold that the 5,000 items actually put forward as the basis of an argument had waived that protection. For the items held back and never used, the refusal to produce was upheld. Records a lawyer uses as a tool in the course of giving legal advice may be handled differently, in other words, but the protection lapses over whatever is actually relied on, and no further.

One attitude runs through all six, Concord included. Courts do not regard chatbot conversations as a specially private domain. Chatbot records are passing now through the place email and workplace messaging passed 20 years ago. BSF's observation that no privacy objection has yet blocked production is a summary of that current.

All of this is a story about American courts. Discovery in this form does not exist in Korea, so the holdings do not cross over as they stand. There is still a reason not to file it away as someone else's problem. Korean outlets have reported investigators examining a suspect's ChatGPT history, and lawyers who now start a matter by asking what the client put to a chatbot. The systems differ, but the condition that the record is sitting somewhere is the same.

The homework these cases hand a company is not a course in doctrine. It is a classification problem. What kind of data the questions employees put to chatbots dozens of times a day are accumulating as, and who can pull them out, is undecided at most companies. Deciding after a suit is filed is too late. By then the other side arrives holding the list.

Three things can be checked today. None of them requires buying a new system.

• • Where work conversations accumulate. On a company license, retention and access are under the company's control. Where the habit of asking work questions from a personal account persists, those records pile up somewhere the company can neither delete nor defend. When a suit calls for them, the company finds out only that it has no control over them
• • That deleting is not the answer. In the Krafton case the deleted logs came back as copies shared over the internal messenger. The larger problem is that the deletion itself becomes an issue. From the point a dispute is reasonably anticipated the rule is not to delete, and who announces that point, and how, has to be settled in advance
• • Whether use for legal advice is kept separate. Heppner held that a chatbot carries no privilege of its own, and Concord left room for work product protection over prompts entered at counsel's direction. That difference can only be argued where the record shows who asked the question and why

All three point to the same place. Not knowing where a record lands and under what label, a company cannot decide which side that data falls on. Chatbot conversations often have no name yet in a company's inventory of data, while courts already count them as ordinary electronic records. What is strange is data being called first in a courtroom and not in the inventory, not the way courts are ruling.

Thanks for reading this far. The facts in this article were checked against the plaintiff's filing as reported by the Volokh Conspiracy, eDiscovery Today's account of the Nevada decision, and BSF's survey of the case law. The Fontainebleau lawsuit is still pending, and when the court fixes the scope of production we will cover that separately. If you have settled on a way of classifying chatbot conversations inside your own company, we would like to hear about it.

Pebblous Data Communication Team
October 20, 2026

• 1.Volokh, E. (2026). "The AI Data Will Provide an Unrivaled View Into the Mind of a Defendant." The Volokh Conspiracy, Reason.com, October 19, 2026. The demand for production in Fontainebleau Florida Hotel, LLC v. Botach (S.D. Fla., No. 1:25-cv-20251) — primary source for Sections 1 and 2.
• 2.Austin, D. (2026). "'Contrary to Law' to Order the Production of the Entirety of Expert's ChatGPT History Log, Court Rules." eDiscovery Today, October 12, 2026. The September 29, 2026 decision in Villanueva v. Las Vegas Metropolitan Police Department (D. Nev., 2:24-cv-00125) — primary source for Section 3.
• 3.Boies Schiller Flexner LLP. (2026). "Are AI Chatbot Prompts Discoverable? Courts Say Yes." BSF Perspectives. A survey of In re OpenAI, United States v. Heppner and Concord Music Group v. Anthropic, among others — source for Sections 4 and 5.