AI, Privilege, and Discovery: Divergent Approaches in United States v. Heppner and Warner v. Gilbarco

Banner artwork by Jade ThaiCatwalk / Shutterstock.com

Two federal courts — one in the Southern District of New York and one in the Eastern District of Michigan — recently confronted similar questions arising from a thoroughly modern problem: how the attorney‑client privilege and work‑product doctrine apply when parties use generative AI during litigation.

Although both cases involved parties’ interactions with AI tools (Anthropic’s Claude in United States v. Heppner and ChatGPT in Warner v. Gilbarco), the courts arrived at opposite outcomes regarding privilege and protection.

For in-house counsel and corporate legal departments, these decisions are more than academic. Employees and business teams are increasingly using generative AI tools in the ordinary course of work, including for tasks that may touch on legal advice, internal investigations, and dispute-related analyses. These cases provide an early signal of how courts may treat AI-assisted communications in discovery — and where privilege protections may break down.

The cases

At a high level, these cases raise two practical questions for legal departments:

1. whether using generative AI tools risks waiving attorney-client privilege or work product protection, and
2. whether the answer depends on who is using the tool, how it is used, and under what terms.

United States v. Heppner (S.D.N.Y. Feb. 17, 2026)

Bradley Heppner, charged with large‑scale securities and wire fraud, created approximately 31 documents memorializing his exchanges with Claude after receiving a grand jury subpoena. His counsel later claimed these “AI Documents” contained defense strategy and were created in anticipation of indictment.

The government seized the documents during a search and sought a ruling that they were not privileged.

Judge Jed S. Rakoff held that the AI Documents were not protected by attorney‑client privilege or work‑product doctrine because they failed to satisfy the core elements of attorney-client privilege: confidentiality, communications between lawyer and client, and having the purpose of obtaining legal advice. Judge Rakoff first noted the absence of an attorney‑client relationship. As explicitly stated in its terms and conditions, Claude is not a lawyer. Because Heppner, a non-lawyer, used the AI tool on his own initiative, despite having legal counsel representing him, the communication did not take place between a lawyer and client.

Second, the court held that there was no reasonable expectation that the communications were confidential. Anthropic’s privacy policy expressly states that user inputs and outputs may be retained, used for training, and disclosed to third parties, including regulatory authorities.

Third, the court also found that Heppner did not consult with Claude or use it to create documents for the purpose of obtaining legal advice, because the terms and conditions of Claude explicitly disclaim its ability to give legal advice. Even if Heppner later shared the AI outputs with counsel, the materials were never privileged at inception and “are not somehow alchemically changed” into privileged documents by later transmission.

Finally, the work product doctrine could not apply because the documents were not prepared by or at the direction of counsel and did not reveal counsel’s mental impressions.

Because Heppner’s self‑directed AI use did not satisfy the elements required to show attorney-client privilege, the government was allowed to inspect the AI documents in full. The court’s reasoning suggests that the absence of counsel involvement — and the presence of third-party platform terms permitting data use — were central to the loss of protection.

Warner v. Gilbarco (E.D. Mich. Feb. 10, 2026)

In this employment‑discrimination case, defendants sought discovery of documents created by the plaintiff using AI tools — particularly internal drafts and analyses she may have generated using ChatGPT. They argued this information could be relevant and that any protection was waived by disclosure to a non‑human third party.

Magistrate Judge Anthony P. Patti denied the defendants’ request, holding that the plaintiff’s AI‑related materials were not discoverable and were protected as work product.

At the outset, the magistrate noted that pro se plaintiffs (those representing themselves without counsel) are entitled to protect their internal mental impressions, analyses, and litigation when preparing documents for their case.

The magistrate then rejected the argument that the use of ChatGPT waived privilege protections because waiver of work product must involve disclosure to an adversary or occur in a manner likely to reach an adversary. Additionally, AI platforms like ChatGPT are “tools, not persons,” and using them does not equate to disclosure to a third party. There is also no evidence suggesting that plaintiff uploaded confidential protected documents into ChatGPT.

Because the plaintiff’s use of ChatGPT to generate documents associated with her litigation were protected by work product privilege, the plaintiff did not have to disclose her AI‑related materials to the defendants.

Contrasting the outcomes

While the underlying facts of these cases vary significantly, some important similarities exist. First, both courts looked to the identity of who engaged with AI to create the purportedly privileged documents. In Heppner, the court noted that Heppner was not a lawyer and was in fact represented by outside counsel. The court invalidated his privilege claim over his use of Claude. Warner was also not a lawyer, but because she represented herself in the case, the court considered her a lawyer in this instance and determined that her use of AI was protected because of her legal status in the litigation.

Both courts also applied time-tested legal theories, while at the same time acknowledging the “new frontier” AI presents to the legal field. The judge in Heppner focused on the core elements of attorney-client privilege — communications between lawyer and client, confidentiality, and legal purpose. In Warner, the magistrate judge focused on the requirement that, in order to waive work product privilege, the document must be disclosed in such a way that an adversary may be able to receive it.

One significant difference is that Heppner focuses on the AI’s terms and conditions, specifically the notice that third parties, including regulators, can access information supplied to and created by the platform. The judge highlighted also that Claude’s terms explicitly state that it is not a lawyer and cannot give legal advice. Warner makes no mention of terms and conditions, even though ChatGPT has similar warnings about third party disclosure.

This distinction may be significant for corporate users. Many AI platforms — including both public and enterprise offerings — contain terms addressing data retention, model training, and potential third-party access. Courts may increasingly look to these provisions when evaluating whether a user had a reasonable expectation of confidentiality. Notably, the Warner case did not address these issues, leaving open how terms of service might affect similar claims in future cases.

These two cases are the first of many involving the intersection of legal privilege and AI. While more cases will be decided in the coming months and years, for now, many questions remain unanswered. Nonetheless, these cases serve as a reminder to in-house counsel to take precautions to prevent the inadequate loss of privilege when using AI.