[Editor’s Note: This article has been republished with permission. It was originally published November 4, 2026 on the Minerva26 Blog]
A Southern District of Florida magistrate judge held that a defamation defendant’s AI prompts and outputs are discoverable because they may show the specific intent to harm that removes Florida’s cap on punitive damages. The final post shows what was said; the prompt history can show why. The lesson: AI prompts are another source of ESI, discoverable when they go to an element of a claim or damages and protected only when they meet the ordinary work product test. Reading time: approximately 11 minutes.
Podcast | Transcript
The Prompt History Is the Process
All year on Case of the Week, we have been tracking how courts handle AI in discovery, and this decision shows where those cases are landing. AI prompts and outputs are just another source of ESI. They are discoverable when they are tied to a claim, a defense, or an element of damages, and they are protected only when they meet the same work product test as any other document.
In Fontainebleau Florida Hotel, LLC v. Botach, the element was intent. The hotel alleges that the defendant used AI tools to help write negative social media posts about the hotel and its lawyers, before the lawsuit and after. It wanted his prompts, his drafts, and the chats he deleted, because the final posts show what he said, but the prompts may show why he said it and what he was trying to accomplish. United States Magistrate Judge Marty Fulgueira Elfenbein agreed that intent is exactly what those prompts can reveal, and she got there without a single new rule for AI.
The Only Issue Left Is Damages
The hotel sued Jacob Shmuel Botach, known publicly as Shmuley Boteach, for defamation and tortious interference over his public statements about an encounter in the hotel lobby in December 2024 and the hotel’s response. The parties sharply dispute what happened, and nothing in this post takes a position on it. Liability has already been resolved against Boteach as a sanction, so discovery was reopened for damages only. That limit matters, because every request had to tie to damages. The docket has moved quickly since this order, with a court-ordered production deadline, an emergency motion to stay that was denied the same day, and a damages trial set for February 2027. The full timeline is in the episode and the transcript.
The hotel is seeking uncapped punitive damages. It served a request for production on Boteach and proposed subpoenas to four AI providers, which the order does not name, seeking his AI communications since December 3, 2024, about the hotel, its counsel, its employees and representatives, the litigation, court orders, and the judges, plus deleted chats. Boteach argued that damages discovery should be limited to his net worth and that general hostility toward the hotel does not prove actual malice.
What Makes Punitive Damages Uncapped in Florida
This is the part of the opinion I want you to focus on. ESI has to be tied to a claim, a defense, or an element of damages to be discoverable, and the Court does an excellent job of doing exactly that, element by element.
The Court started by rejecting the idea that punitive damages discovery is only about net worth. It looked to the Supreme Court’s decision in State Farm v. Campbell, which called the degree of reprehensibility the “most important” consideration in evaluating punitive damages, including whether the conduct was repeated rather than isolated and whether the harm resulted from intentional malice, trickery, or deceit rather than accident.
Then the Court turned to Florida law, which independently makes intent central. Section 768.72(2) of the Florida Statutes distinguishes intentional misconduct from gross negligence, and section 768.73(1) sets different caps depending on the nature of the conduct. Section 768.73(1)(c) removes the cap entirely: “Where the factfinder determines that at the time of injury the defendant had a specific intent to harm the claimant and determines that the defendant’s conduct did in fact harm the claimant, there shall be no cap on punitive damages.” Looking to Florida Supreme Court precedent holding that punitive damages exist to punish and deter, and that the wrongdoer’s intent informs the inquiry, the Court concluded that Boteach’s alleged specific intent to injure the hotel is not a collateral matter. It bears directly on the damages the hotel seeks.
That is why this matters beyond Florida. The relevance of these chats comes from a specific statutory element: specific intent to harm as the trigger for uncapped punitive damages. In a jurisdiction without that framework, or in a case where intent is not tied to a remaining issue, the same request could come out very differently. The AI chats are not discoverable because they are AI. They are discoverable because they are ESI that goes to an element of damages that is still open.
Why the Prompts Go to Intent
The Court connected the prompts to that element through the Supreme Court’s decision in Herbert v. Lando, which refused to shield a defamation defendant’s editorial thought processes from discovery when his state of mind bears on a legal issue. Then it explained why prompts are different from the posts themselves: “To the extent Defendant used AI platforms to generate, revise, or refine statements concerning Plaintiff, its counsel, or its representatives, those communications may reveal not merely the final content of Defendant’s statements but also the purpose for which the statements were created, the instructions Defendant supplied, and the effects he sought to achieve.”
Think about what that means in practice. The public post is the end product; the prompt history is the process of getting there. A prompt history can show whether someone asked a tool to summarize what happened or asked it to make a post more damaging. It can show drafts that were softened and then made harsher, instructions about which audience to reach, or the same theme worked over and over again across months. None of that is visible in the final post, and all of it could bear on whether someone specifically intended to cause harm and whether the conduct was repeated rather than isolated. I am not saying any of that exists here, because we have not seen these chats, but that is what prompt-history discovery can reveal.
The Court was careful about the limits. It did not find that the chats will establish malice or specific intent, only that they bear a sufficient relationship to those disputed damages issues to warrant discovery. It also rejected Boteach’s actual malice argument, and the distinction is worth understanding. Constitutional actual malice is about knowledge of falsity or reckless disregard for the truth; reprehensibility and the specific intent exception are about culpability and intent to harm. As the Court put it, the hotel seeks evidence “not simply to establish that Defendant disliked the Hotel, but to determine whether Defendant deliberately engaged in conduct intended to cause injury.” And chats created after the original posts are fair game, because later statements can be circumstantial evidence of earlier intent or of a continuing course of conduct. Timing goes to weight, not discoverability.
Where the Court Drew the Line
Because intent was the hook, the Court allowed only the categories aimed at the people the alleged harm targeted: AI communications about the hotel, its counsel, and its employees, owners, or representatives, including deleted chats in those categories. It cut the categories that swept in everything. “This litigation” was exceptionally broad, chats about court orders may just reflect Boteach trying to understand an order, and for the judges the question was relevance to damages, not whether he criticized the judiciary. A footnote makes clear that a responsive chat is still produced even if it also mentions the litigation, an order, or a judge.
Intent Evidence Is Not Work Product Just Because AI Was Involved
Boteach argued the chats were work product. The Court read Rule 26(b)(3)(A) to cover materials a party prepares itself, not just lawyers, but only when the primary motivating purpose was to aid in the litigation, and the party asserting protection has to prove it. Reconciling Heppner and Warner from Episode 183, it held that neither case creates a categorical rule. Chats Boteach created to produce public statements about the hotel are not work product, because their purpose was public-facing content, not his defense; anything genuinely prepared because of the litigation can be withheld if he logs it.
The line that sums up this episode: “a party’s communications with and use of AI does not automatically establish either discoverability or protection. The analysis turns on the ordinary requirements of Rules 26(b)(1), 26(b)(3), and 26(b)(5).” That is the same analysis you would run on email or text messages. AI did not change the rules.
Getting the Prompts He Deleted
The hotel went to the providers because it was concerned Boteach may have deleted or failed to preserve responsive chats. The Court allowed the subpoenas but refused to compel Boteach to sign consents authorizing the providers to release his records, holding that neither Rule 34 nor Rule 45 requires a party to authorize a third party to hand its records to an adversary. It also emphasized that permitting the subpoenas does not establish that they are enforceable. My read is that the Stored Communications Act will be the obstacle. We covered it in Episode 167, Sihler v. Microsoft. The SCA generally bars covered providers from disclosing communications content in response to civil process without consent. I have not seen a court decide whether an AI provider is covered, but given how the SCA has been interpreted, I think it very likely is, which would make the consent this Court refused to compel the hotel’s only path to those chats.
The order also says nothing about when any chats were deleted, and that is the question I want answered. If Boteach deleted chats after his duty to preserve attached, at the latest when suit was filed in January 2025, there is a potential Rule 37(e) issue. Rule 37(e) applies only when ESI cannot be restored or replaced through additional discovery, and a provider subpoena is exactly that additional discovery. If the provider still has the chats, there may be no loss at all; if not, the hotel will have a record of what was lost and when.
Who Gets the Provider Production First
Normally a nonparty produces to the requesting party, which shares a copy with the other side. But if the providers produce directly to the hotel’s counsel, Boteach never gets a chance to screen for work product and log it, so the Court directed the parties to meet and confer on how the providers will produce. Neither of the later orders tells us how that was resolved, and it is complicated. Boteach likely still has some of these chats, and he knows which ones were deleted. There is no evidence he deleted anything, and I am not suggesting he did, but the practical realities make this a hard production to negotiate. If I were advising either side, these are the three options I would put on the table.
Option one: defense counsel reviews first. The providers produce to Boteach’s counsel, who reviews on a short court-ordered deadline, produces everything nonprivileged, and logs anything withheld. To make sure nothing disappears in between, the providers also give the hotel a count or a content-free index of what they produced, so the two productions can be reconciled. Think metadata fields about each chat without the content. That would let the hotel line up the dates against the social posts that went out about it, build a timeline of what transpired, and see where any deleted chats fit.
Option two: a neutral screens. The providers produce to a neutral vendor or special master, who screens with defense counsel and releases the nonprivileged material. It is the most credible process, but it costs more and takes longer, and timing is the issue with a trial set for February.
Option three: simultaneous production under a clawback order. The providers produce to both sides at once under a Rule 502(d) order. It is the fastest, but once the hotel’s counsel has seen protected material, a clawback cannot undo that. The protective order’s 30-day Outside Counsel Eyes Only treatment for nonparty productions does not help either, because outside counsel is exactly who should not see work product. You cannot unring the bell.
All of this has to fit a schedule where damages discovery closes November 9 and trial starts February 22, 2027, and the providers need time to respond and may object. That makes option one, with a short review window and a reconciliation step, the most realistic.
What to Do This Week
Tie every AI request to an element. The categories that survived here were tied to the statutory trigger for uncapped punitive damages, specific intent to harm, and the ones that failed swept in everything. That is consistent with what we have been saying about discovery on Case of the Week for close to five years. When you draft requests for AI prompts, start with the element you need to prove and work backward to the subjects, the date range, and the people.
Know what your client has been doing with AI, and know it early. Put AI accounts on your intake checklist and in your legal hold. The consumer policies I checked for two major providers, OpenAI and Anthropic, say deleted chats are removed within 30 days, subject to legal exceptions; a data processing agreement for a non-consumer account may differ. Deletion after the duty to preserve attaches is a spoliation problem. If you are seeking the data, send a preservation letter to the other side and to the provider the day you identify the account. Start early.
Treat work product for AI use as earned chat by chat. If your client is going to use AI on the case, have them do it in a separate account or project, ideally at your direction, away from whatever they use for public statements. Document, document, document. The more you know about the AI use, the better.
Don’t count on a court compelling consent. Plan for a provider that will not produce without it, think through the Stored Communications Act before you serve, and build the production protocol into the meet and confer before the subpoena goes out.
What to Watch Next
ECF No. 289 is a magistrate judge’s nondispositive order, reviewable under Rule 72(a) and Rule 4(a)(1) of the Southern District of Florida’s Magistrate Judge Rules within 14 days of service, for clear error or a result contrary to law. Boteach’s emergency motion to stay his October 30 production deadline was denied the same day, and as of our November 3 recording, no objection or appeal had been filed. Watch whether the providers object, whether anyone raises the Stored Communications Act, and whether the provider productions show chats deleted after the duty to preserve attached before the February 2027 damages trial.
She is a co-author of Electronic Discovery and Records and Information Management, and host of Case of the Week at Minerva26. As Principal at ESI Attorneys, Kelly manages the boutique eDiscovery and information law firm that acts as operational business partners with its clients to advise law firms, corporations, and municipalities on all areas of electronic information including eDiscovery, privacy, cybersecurity, and information governance.
Kelly is also the CEO of Minerva26. — a SaaS-based practical resource for litigators handling eDiscovery — that curates discovery decisions, rules, and additional content. She is developing an online academy to provide on-demand education for lawyers and legal support professionals to stay abreast of changes in the law and technology that affect litigation and clients’ obligations to respond.
You can reach Kelly at [email protected], join her Facebook community group at Let’s Talk eDiscovery, or connect with her on Twitter @kellytwigger.