Four decisions in nine months, and they do not agree. One court found a defendant’s AI conversations unprotected; three have found work product protection, and one of those quashed a subpoena to OpenAI outright.
Whether a party’s own conversations with a generative AI tool are discoverable is now a live question with actual decisions behind it. They point in different directions, and the direction depends on facts that are easy to state.
The case that found no protection
In United States v. Heppner, No. 1:25-cr-00503 (JSR) (S.D.N.Y.), Judge Jed S. Rakoff ruled from the bench on 10 February 2026 and issued a written opinion on 17 February 2026, holding that a defendant’s exchanges with a generative AI platform were protected by neither the attorney-client privilege nor the work product doctrine.
As reported, the privilege analysis failed on more than one ground: the platform is not an attorney; the exchanges were not confidential given a provider privacy policy permitting collection, use for training and disclosure to third parties; and the defendant was not using the tool to obtain legal advice. The work product claim failed because the materials were not prepared by or at the behest of counsel and did not reflect counsel’s strategy.
Two facts about this decision matter more than the holding. It is a criminal case, and the defendant was represented by counsel. Every court that has since declined to follow it has said so on those grounds.
The three that found protection
In Warner v. Gilbarco, Inc., No. 2:24-cv-12333 (E.D. Mich.), Magistrate Judge Anthony P. Patti held on 10 February 2026 that a self-represented plaintiff’s AI-assisted materials were protected work product. The reasoning turns on the text of the rule: Federal Rule of Civil Procedure 26(b)(3)(A) protects materials prepared “by or for another party or its representative,” not only materials prepared by or at the direction of an attorney. Using a public AI platform did not waive the protection, and the request was treated as a fishing expedition into internal analysis.
In Morgan v. V2X, Inc., No. 25-cv-01991-SKC-MDB (D. Colo.), Magistrate Judge Dominguez Braswell held on 30 March 2026 that “Rule 26(b)(3) applies and offers Plaintiff some level of work product protection in connection with his use of AI,” and that using AI does not automatically waive protection. The court distinguished Heppner as a criminal case involving a represented defendant and cited Warner as consistent.
And in Assini v. Hayward, 2026 NY Slip Op 26086 (Sup. Ct., Nassau County, 4 June 2026), Justice Fischer quashed a subpoena and amended subpoena served on OpenAI OpCo, LLC seeking a self-represented defendant’s ChatGPT prompts and outputs, holding that OpenAI “is not required to provide records in response to the Subpoenas unless a further court order should issue.” The court expressly adopted Morgan over Heppner, holding Heppner non-binding because it was criminal, and quoted Morgan’s observation that “the use of AI closely resembles the kind of confidential, strategy-laden iterative work product” and that AI use “does not eliminate all expectations of privacy or automatically waive protections.”
Where the line appears to be forming
On four decisions it is early, and the pattern is worth stating as a pattern rather than a rule:
- Self-represented parties have fared better than represented ones, because Rule 26(b)(3)(A) protects a party’s own preparation and does not require an attorney’s hand
- The criminal and civil contexts are being treated differently, and three courts have said so
- Use of a public, consumer-grade platform has not by itself waived protection in the civil cases — but it was part of why confidentiality failed in the criminal one
- No court has held that AI interactions are categorically protected, and none has held they are categorically discoverable
The provider is not a safe harbor, and one order says what to do about it
Morgan contains the most consequential practical holding in the set. The court required the plaintiff to disclose which AI platform he had used for confidential information, reasoning that the opposing party is entitled to know which system received it.
It then amended the protective order on its own motion to bar inputting confidential information into any AI platform unless the provider is contractually barred from training on inputs and from unrestricted third-party disclosure, and is obliged to delete on request. The court acknowledged that this standard excludes most mainstream low-cost tools.
That is the operative lesson for anyone negotiating a protective order now: the terms that decide this are the provider’s contract terms, not the party’s intentions.
Separately: the platform itself can be ordered to produce at scale
The privilege question is about a party’s own logs. A different question is whether the provider can be made to produce logs in bulk, and there is an answer.
In In re: OpenAI, Inc., Copyright Infringement Litigation, No. 25-md-3143 (SHS) (OTW) (S.D.N.Y.), Magistrate Judge Ona T. Wang granted a motion to compel on 7 November 2025 at Dkt. No. 734, ordering production of 20 million de-identified consumer ChatGPT conversation logs — a sample drawn from tens of billions of retained logs. The news plaintiffs had originally sought 120 million.
The entire sample had to be produced, including logs containing none of the plaintiffs’ works, because they bear on the fair use defense. OpenAI’s alternative proposal to run search terms and produce only matching conversations was rejected. Judge Sidney H. Stein affirmed on 5 January 2026 at Dkt. No. 1021, denying OpenAI’s objections.
Be precise about what that order is and is not. It is 20 million de-identified consumer logs produced under a protective order as a sample, not every log OpenAI holds, and not the identified conversations of named individuals.
What this means operationally
Four things follow for anyone running discovery in a matter where AI use is plausible, which is now most matters.
Preservation. If a party used a generative tool in connection with the dispute, the interactions are potentially discoverable material subject to the same hold analysis as anything else. Provider retention settings are a configuration, and some platforms default to deleting after a period, so the preservation question arrives early and has a clock on it.
Collection. Export capability differs by platform and by account tier, and an individual consumer account, an enterprise tenant and an API account expose different things. The practical question is the same one as for any cloud source: what does the export actually produce, and what does it leave behind.
Protective order drafting. Morgan turned on provider contract terms. A protective order drafted today without an AI provision is a protective order that will be amended later, possibly by the court on its own motion.
Privilege review. No court has addressed the use of a generative tool to perform privilege review, which is a different question again. What these four cases address is AI interactions as evidence.
One caution about the state of this
Four decisions, three of them by magistrate judges and one by a state trial court, over nine months. Nothing here is settled, Heppner has been distinguished three times but not overruled, and the next decision may reorder the whole picture.
Anything written about this subject — including this article — should be checked against what has happened since. Confirm the current state before relying on any of it.
What we do
Pro-Legal Services, Inc. supports discovery involving generative AI interactions from Arlington, Virginia — preservation scoped against the platform’s actual retention settings, collection against what a given account tier will export, and the collection record documented at the time rather than reconstructed.
Small business, minority-owned, SAM registered. Collection, review and production in one building.
Pro-Legal Services, Inc. supports litigation teams from Arlington, Virginia, for matters in federal and state courts nationwide.
AI use in the matter and a hold going out? Tell us the platforms and the account types. What is exportable depends on the tier, and retention settings run on their own clock. Pro-Legal Services, Inc. · 1550 Wilson Blvd, Suite 250, Arlington, VA 22209 · 202-955-9400 · info@pro-legal.com