A Miami judge let a hotel go after a rabbi’s AI chats in October. When he didn’t comply, she ordered him to authorize the release of all his account data and content.
On October 21, 2026, a federal magistrate judge in Miami signed a 25-page discovery order that I think every litigator should read. Judge Marty Fulgueira Elfenbein let the Fontainebleau Miami Beach go after Rabbi Shmuley Boteach’s conversations with AI chatbots. The hotel sued him for defamation, and it wants his prompts because it thinks they’ll show what he was trying to do.
The headlines made it sound like the hotel gets to read his whole AI history. That overstated the October order, which was narrower than the coverage and more useful. I think it’s the most practical guide so far to how a court will sort a litigant’s AI chats. Judge Elfenbein sorted them by subject and by purpose, and she made the person claiming protection prove it one chat at a time.
Then the case kept moving. In November, the judge ordered Boteach to sign a consent so his AI providers could disclose deletion logs. He edited the form and sent it through his lawyers, and the providers couldn’t process it. Five days later, she ordered him to sign a much broader consent covering all his account data and content. That second stage may teach your clients more than the first. It shows how fast a narrow order can widen once a party stops following it.
A quick note before going further. I’m not a lawyer, and nothing here is legal advice. I spend my days helping law firms figure out how they and their clients should use AI, and these orders change what I’d tell both.
How a lobby argument became an AI discovery fight
The two sides tell the story differently, so I’ll stick to the court record. In early December 2024, another guest confronted Boteach in the Fontainebleau lobby over the war in Gaza. Boteach then went public. He accused the hotel of tolerating antisemitism and told his followers to stay away. The Real Deal put his following at 1.2 million across platforms at the time. He also announced a $100 million lawsuit of his own. The hotel sued him for defamation and tortious interference, in a case that ended up in federal court, and its complaint alleged he used AI to make videos of people tied to the property.
It went badly for him from there. The court had already sanctioned him $15,000 for reposting surveillance video covered by a protective order. In February 2026, the court told him to stop posting about the hotel and its lawyers. He kept posting anyway. On September 17, 2026, the court struck his defenses and entered a default against him on liability. He lost that half of the case as a sanction, without a trial on the merits.
That left only one question, which is how much he pays.
Why the prompt matters more than the post
Florida law generally caps punitive damages at three times compensatory damages or $500,000, whichever is greater. The statute has exceptions, including a higher cap for some conduct driven by financial gain. The one that matters here removes the cap if the jury finds the defendant had a specific intent to harm at the time of the injury and actually caused harm. The hotel told the court it wants uncapped punitive damages, and it has said the award could top $20 million. Once the cap turns on intent, his state of mind becomes the most valuable fact left in the case.
The hotel’s theory was simple. His public posts are already in the record. What it wanted was the instructions behind them, because those might show why he wrote what he did. Judge Elfenbein agreed that his AI chats could reveal “the purpose for which the statements were created, the instructions Defendant supplied, and the effects he sought to achieve.”
I think that sentence will show up in a lot of document requests next year. Picture how people use these tools when they’re angry. Someone might type “make this sound meaner” or “write a post that will hurt their bookings.” Those are made-up examples, but you get the idea. A finished post leaves room to argue about tone and meaning, and a prompt asking for maximum damage leaves almost none.
The judge was careful about how far she went. She said she wasn’t deciding that the chats would prove malice or intent to harm, only that the hotel could look. Boteach had argued that damages discovery should stop at his net worth, and she rejected that too. Under the Supreme Court’s State Farm decision, how blameworthy a defendant was also counts.
What the hotel got, and what it didn’t
The hotel asked for every AI chat since December 3, 2024 on six subjects, plus any deleted chats on those subjects. It won half.
It can get his chats about the hotel and anyone who works for it or represents it, including the lawyers suing him. Deleted chats in those categories count too. Those requests were tied to the people connected to the harm, and that’s why they survived.
The judge refused the requests aimed only at the case itself. She wouldn’t make him produce every chat about the lawsuit. The same went for chats about her orders and about Judge K. Michael Moore or Judge Elfenbein herself. She called the request for everything about the litigation “exceptionally broad.”
But footnote 4 closes a loophole, and lawyers on both sides should read it. A chat that falls in an allowed category can’t be held back just because it also mentions the litigation, a court order or a judge. It stays subject to production unless a valid privilege or protection applies. So a chat about the hotel that also works through Boteach’s defense is still in play, and he’d need a real work-product claim to keep it.
For your clients, the subject of a chat decides whether it’s in the request at all, and its purpose decides whether it can be withheld.
The ruling lawyers should read twice
The early AI privilege decisions of 2026 pointed in different directions.
In February, Judge Jed Rakoff in New York ruled in United States v. Heppner that about 31 documents a criminal defendant made with the consumer version of Claude weren’t privileged and weren’t work product. Heppner had used the tool on his own initiative, before any lawyer directed him to. Rakoff said a chatbot isn’t a lawyer. He also pointed to Anthropic’s privacy policy to find the chats weren’t confidential.
That same month, a magistrate judge in Michigan went the other way in Warner v. Gilbarco. A former employee suing without a lawyer had used ChatGPT to help draft court filings. The court protected those chats as work product. It also held that typing your work into an AI tool isn’t the same as handing it to the other side.
Other courts followed Warner through the spring. Morgan v. V2X in Colorado (March 30) and Tym v. Cerno in New Mexico (April 22) both said AI use doesn’t automatically destroy work-product protection. Each of those cases involved someone representing themselves. In June, the Texas Business Court applied the same idea to a represented company in Tate Group Automotive v. Legacy Automotive Capital, and it declined to follow Heppner.
Judge Elfenbein’s order joins that line. She held that a represented party’s own AI chats can be protected work product even when no lawyer directed them. Then she made that protection hard to get. Boteach has to show, chat by chat and with evidence, that the main reason he created each one was to help with the case. His blanket objection failed. Anything he holds back goes on a privilege log the hotel can attack.
The sentence I’d remember from her order is a short one. Using AI “does not automatically establish either discoverability or protection.” What decides it is why the chat was created.
She also said something every client should hear. Chats used to write or spread content about the other side “ordinarily would not satisfy” the work-product rule. I suspect that’s the most common way people use a chatbot during a fight. They draft the post or the email they’re about to send.
What the October order doesn’t say
The coverage blurred a few points, so I want to get them right.
The order reaches only Boteach’s own chats. It doesn’t touch the hotel’s AI use or his lawyers’ AI use. When it mentions chats about “counsel,” it means the hotel’s lawyers. He had posted about one of them by name, which helps explain why that category survived.
The October order never identifies the four AI companies involved. The November orders name two of his accounts, ChatGPT and Grok. Neither the October order nor the November 11 order separates consumer accounts from enterprise accounts, and neither turns on terms of service. That’s a real difference from Heppner, where the privacy policy did a lot of the work.
The October order also never analyzes attorney-client privilege. Every argument in it is about work product. That leaves the hardest case open. Think of a client who pastes the lawyer’s advice into a chatbot during a rant about the other side. Heppner suggests that chat isn’t privileged. Nothing in these orders says otherwise, and I’d tell clients to act as if Heppner is right.
Where the subpoenas stood in October
In October, the judge let the hotel serve subpoenas on four AI providers. She also refused to make Boteach sign releases authorizing those companies to hand over his chats. The hotel hadn’t cited any authority for forcing him to.
That refusal mattered because of a 1986 federal law, the Stored Communications Act. It generally bars these kinds of providers from disclosing the contents of stored messages in response to a civil subpoena. User consent is the usual way around it. A Colorado Lawyer article from this spring argues the law likely blocks civil subpoenas to AI companies for chat content, though no court has squarely decided the question. The same article points out that some courts have ordered a party to give that consent.
So on October 21, the practical road to his chats ran through Boteach himself. That changed within three weeks.
What happened next
The October order wasn’t the end of it. On November 6, in an order docketed as ECF 322, Judge Elfenbein told Boteach to finish producing his nonprivileged AI chats and to log anything he withheld. She also ordered him to sign a written consent letting his AI providers disclose deletion logs covering a two-year period. That consent covered deletion logs only, and she made no finding that he had deleted anything.
He signed it, but not the way the court asked. According to the November 11 order, the hotel told the court he had changed the form’s wording and left out his account details. He also skipped the notarization and sent the form through his lawyers instead of from his AI account email addresses. The hotel’s lawyer read into the record an email in which ChatGPT’s in-house counsel laid out what was wrong with it. As of the November 9 hearing, he still hadn’t produced a single responsive AI chat. His account data was sitting with an e-discovery vendor, and his own lawyers didn’t have access to it yet.
The judge found that his edits made the form useless, and she went further. The November 11 order requires new consents, signed without changing a word, that authorize the providers to disclose all account data and content, deletion logs included. It names ChatGPT/OpenAI and Grok/xAI. He had to fill in every account detail and get the forms notarized. The deadline to send them was 9 a.m. on November 12, from the email address tied to each account, plus xAI’s privacy portal for Grok.
She also ordered him to pay the hotel’s reasonable fees for the November 5 and November 9 hearings under Rule 37. She found he had failed to follow at least two discovery orders. And she wanted a knowledgeable person from his e-discovery vendor at the November 12 status conference to answer her questions.
That doesn’t mean the hotel gets to read everything. The order has Boteach’s side review what comes back. If the deletion logs show nothing was deleted, no second review is needed. If they show deletions, he has to review the new data and produce any responsive, nonprivileged chats he hadn’t already turned over. The judge also took under advisement the hotel’s request to find that he waived privilege by missing the deadline for his privilege log.
The sequence matters more than any single ruling. I think a party who kept up with the October order could have stayed inside the narrow lines it drew. Once compliance slipped, the fight moved from what the chats say to whether he kept them and turned them over. The judge put it bluntly when she wrote that a party “may not selectively comply with an order.”
Deleted chats are where these fights get expensive. In Miller v. Regions Bank, a lawyer deleted his ChatGPT account three days after a federal judge in Alabama ordered him to produce its history. On May 21, 2026, the court inferred the missing history would have shown ChatGPT wrote the fake quotations in his brief. It also suspended him from practicing in that court for six months and referred him to the state bar. Neither November order finds that Boteach deleted anything, but both show the court is ready to check.
What this changes for your clients
Start with the limits. These are one magistrate judge’s interim orders, in a case where the defendant had already been sanctioned into a default. They bind no other court. A judge in an ordinary commercial dispute may be less generous to the side asking.
Still, I expect other courts to borrow the approach. Any case where intent drives the outcome is now a candidate for an AI-chat request. That includes punitive damages claims, insurance bad faith, employment retaliation and contested custody.
For the lawyer asking, the lesson is to tie every request to one element and a start date. Half the hotel’s requests died because the judge found them too far removed from damages. For the lawyer defending, blanket objections didn’t work here, and I wouldn’t count on them anywhere. Somebody has to read each chat and log the ones worth fighting for, and that’s real associate time. Get the log in on time, too. Missing that deadline turned into an open waiver question in this case.
The November orders teach something more basic. If a court orders a consent form, sign it as written and send it the way the order says. Boteach’s edits turned a narrow deletion-log consent into a full-account authorization in five days. And if a vendor is processing a client’s AI data, be ready to explain where things stand. This judge wanted someone from the vendor in the room.
The bigger change belongs in the intake meeting. I don’t think most clients see their chatbot history as evidence yet. I’d want every litigation client to hear four things on day one.
Assume anything you type about the other side can end up in front of the other side.
Don’t use AI to draft posts or emails about your opponent while the case is open.
Keep anything that might be relevant, including chats in ChatGPT, Claude, Gemini, Copilot, Grok and any image or video tool you’ve used about this dispute.
If you want AI help preparing, ask us first, and we’ll set it up in writing.
That last point ties the rulings together. Heppner hinted that a lawyer’s direction might change the result, and the October order rewards proof of purpose. A separate project used for litigation prep, with written instructions from counsel and a record of why it exists, may help support a work-product claim. It doesn’t guarantee protection, and the provider’s confidentiality terms still matter.
What to do Monday
Add AI accounts to every litigation hold letter and intake checklist. Name the tools, because “electronically stored information” won’t make a client think of ChatGPT.
Send every active litigation client a short AI warning built on the four points above, and ask each one which AI tools they’ve used about the dispute.
Pull your open matters where intent drives the money and decide whether to serve a narrow AI-chat request. If you’re on the receiving end of one, calendar the privilege log and any consent deadlines like court deadlines.
I think any litigation hold that skips AI accounts is already out of date.
Thanks for reading, I hope you enjoyed this. I saw this case making the rounds, and a few social media posts were a bit misleading about what it decided and what it means for firms and clients. I wanted to give my take.


