How organizations and individuals use generative artificial intelligence (GenAI) tools can create unique privilege issues that parties must then contend with in discovery disputes across the country. Decisions about how to implement those tools, the policies that govern these tools, whether the AI tool is confidential and the level of counsel’s involvement in developing or directing GenAI prompts and outputs may all be important considerations in a court’s determination on whether GenAI-generated outputs are entitled to privilege protection.

Most recently, the US District Court for the Southern District of New York examined this issue in United States v. Heppner. In that case, Judge Jed S. Rakoff examined whether materials the defendant generated using GenAI and shared with defendant’s counsel were protected by the attorney-client privilege or work-product doctrine. The court held these materials were not protected.1

This article discusses the Heppner GenAI privilege ruling, some pertinent considerations in light of other recent GenAI decisions and key takeaways for how organizations can implement GenAI tools in their organization in a manner that safeguards privilege and minimizes waiver risks while still empowering legal counsel to leverage GenAI in legal investigations where appropriate.

The United States v. Heppner decision and the use of GenAI to generate purportedly privileged materials

In United States v. Heppner, the government alleged a Texas executive committed securities fraud and wire fraud through certain false representations that ultimately defrauded investors of over US$150 million. During the discovery phase, the government asked the court to find that approximately 31 documents generated by the defendant executive through the GenAI tool Claude, developed by Anthropic, were not protected by attorney‑client privilege or the work-product doctrine and should be produced.

Importantly, the disputed documents were created by the defendant prior to his arrest using prompts entered into the publicly available version of Claude. Without direction from counsel, the defendant used Claude to outline a potential defense strategy in anticipation of a potential indictment, and later shared those documents with his counsel. According to his counsel, defendant input information he had learned from communications with counsel and had used Claude for the purpose of speaking with counsel to obtain legal advice. Pursuant to the parties’ privilege protocol, the defendant logged those documents as privileged.

After initially ruling from the bench at a pretrial conference that defendant’s Claude outputs were not protected from government inspection by the attorney-client privilege or work-product doctrine, Judge Rakoff set forth the reasoning for the court’s ruling in a subsequent memorandum by examining each of the elements of the attorney-client privilege in turn:

  1. a communication between a client and lawyer
  2. that is intended to be and is kept confidential and
  3. is made for the purpose of obtaining or providing legal advice

He then assessed whether the work-product doctrine applied.

First, the court determined that the defendant’s prompts and Claude’s outputs were not communications between a client and his lawyer because Claude is not a lawyer and the defendant’s use of Claude is akin to a discussion between two non-lawyers, or asking a friend for legal advice. Judge Rakoff noted that the use of Claude in this context was distinct from drafting confidential notes with the intent to share them with a lawyer later because the AI tool was public (specifically, it was not a closed-off enterprise tool) and thus not confidential. Moreover, the court found that AI tools like Claude cannot themselves establish privilege in the absence of an attorney-client relationship; in other words, there is no ‘AI privilege’ between the user and the AI tool. On this issue the court appears to imply throughout the opinion that because the defendant had no reasonable expectation of privacy in his use of Claude, the analysis should focus on whether in choosing to use Claude of his own volition the defendant intended to seek legal advice from Claude, which is not a lawyer and cannot establish its own AI privilege.

Second, the court also determined that neither the prompts nor the outputs were confidential, as the defendant had used a publicly accessible AI platform. According to Anthropic’s privacy policy, Anthropic reserves the right to monitor and disclose any prompts and outputs, including to third parties such as the government. It is perhaps unsurprising that both the government’s motion and the court’s ruling cites to In re: OpenAI Inc. Copyright Infringement Litigation in support of this finding,2 which we have written about previously to discuss several notable opinions on GenAI in discovery issued by the OpenAI court (particularly around preservation obligations and balancing the right to discovery and user’s privacy interests). Here, the court cited an OpenAI decision to conclude that AI users do not have a substantial privacy interest in conversations on a publicly accessible AI platform, because prompts are voluntarily disclosed to the AI platform, and thus the defendant could have no reasonable expectation of confidentiality in his use of Claude.

Third, the court concluded that the communications with Claude were also not made for the purpose of obtaining legal advice, as defendant did not use Claude at the suggestion of or at the direction of counsel, and Claude could not be said to be acting as the lawyer’s agent within the protection of the attorney-client privilege. This said, the court acknowledged this issue was a closer call given the defense’s assertion that the materials were expressly generated for the purpose of speaking with counsel and incorporated information conveyed by counsel to the client. But the court placed greater weight on the fact that the defendant used Claude of his own volition and that, regardless of the underlying purpose or direction, the defendant’s use of the tool was not confidential and defendant had no reasonable expectation in its confidentiality and so any potential privilege that may have existed was waived.

Similarly, the court concluded the AI materials were not protected by the work-product doctrine as they were not prepared by or at the behest of counsel, and did not disclose counsel’s litigation strategy at the time.

Our take on Heppner

It is interesting here that the Heppner court finds no privilege exists with respect to the Claude materials, rather than concluding that privilege exists but has been waived in light of the use of a publicly available AI tool. The court appears to weigh lack of confidentiality heavily in its analysis of each element of the attorney-client privilege, suggesting that because defendant had no reasonable expectation of confidentiality in his use of Claude, the defendant could not be said to have prepared his inputs or AI-generated outputs in the attorney-client context or for the purpose of seeking legal advice. If the tool was actually confidential, the court may have considered these factors more favorably towards a finding of privilege.

Putting aside the issue of confidentiality, the court does not directly address why the use of Claude should be treated differently from the use of any other technology for the purposes of assessing the existence of an attorney-client relationship, particularly where defendant intended, and in fact did, share his AI-generated notes with his counsel in order to obtain legal advice. Practically speaking, a client could use any number of tools to prepare notes in preparation for a conversation with counsel, whether by opening up a word processing document (and using spellcheck), or preparing handwritten notes. In these circumstances, it seems that the relevant analysis would not be whether a privilege exists between the client and the chosen tool, but rather whether the notes were specifically prepared by the client for their lawyer to obtain legal advice, which would not be unique to any particular technology. Depending on the facts of the case, notes prepared by a client in preparation for discussion with their lawyer in order to seek legal advice could arguably be considered privileged regardless of whether those notes are actually shared with or prepared at the direction of counsel.

Furthermore, the Heppner decision left open the question of whether the decision to use an AI tool would have any impact on the applicability of privilege where counsel had directed defendant to prepare the materials. Each situation, and the resulting decision, would likely be highly context-dependent and thus a blanket rule that using AI tools – or any technology – would mean that privilege does not apply would be highly suspect.

Although this case involves an individual executive facing criminal charges, this decision still has implications for companies managing their civil legal matters; whether arising in the criminal or civil context, the overarching tests, standards and analysis courts use to determine whether privilege exists are the same. Perhaps most important, companies should consider how they monitor their employees’ use of publicly available GenAI for business purposes; companies should presume such use can break confidentiality and breach privilege, regardless of whether it is done at the direction of counsel. There are also implications even where employees are directed to conduct business exclusively on enterprise-implemented AI tools (which AI companies typically design to be closed to the public and restricted to the enterprise environment, and in which the AI company that owns the tool does not retain any rights to monitor or disclose). And just like any other technology (for instance, like email), the inputs and outputs from those “closed” AI tools may still be subject to privilege challenges even if there is no dispute that the tool isn’t public, like when non-lawyers prepare materials and there is a question as to whether there is sufficient evidence to demonstrate that the materials were developed for the purpose of seeking legal advice or whether they were created at the direction of counsel. Fortunately, as discussed below, there are steps companies can take to protect privileged information even when using  GenAI.

Other recent GenAI privilege rulings

Heppner is not the only GenAI privilege decision that companies should be aware of though, and they should be aware that not all the decisions are coming out the same way. In contrast to the Heppner court, the court in Concord Music Group v. Anthropic, found that GenAI prompts and outputs from Claude generated by plaintiffs during a pre-suit investigation into potential infringement by Anthropic were protected by the work-product doctrine and did not have to be produced.3

As background, plaintiffs – a group of music publishers – had already produced nearly 5,000 prompts and outputs that supported their claims of infringement, but had objected to production of prompts and outputs that they were not relying on to support their claims that Claude generates infringing content. Plaintiffs argued that those inputs and outputs were privileged.

The court reasoned that, because the queries and prompts were crafted by counsel and contained counsel’s mental impressions, they were not fact work-product, and thus these prompts were squarely protected by the work-product privilege. The court did note that questions remained regarding whether there was potential limited work-product waiver on the basis of the sword-and-shield doctrine, but that did not change the overall analysis. The court also noted that this ruling aligned with existing precedent, finding that similar use of other AI tools like ChatGPT by counsel to support legal investigations had been found to be protected by the work-product doctrine.

Key considerations for companies using GenAI tools

Recent GenAI rulings on privilege serve as a guide to companies that are considering implementing GenAI tools and instructing employees on how to use them. As more and more courts confront these issues, trends will become clearer and precedent will establish a set road for companies to follow. For now though, companies must rely on the decisions that are available to decide how to implement these tools. This implementation is more than just policies and procedures, and requires thoughtful consideration of which GenAI tools are permissible for business use and how they are used, providing adequate guidance to employees on appropriate use in legal matters, monitoring employee compliance, understanding the pitfalls with how employees actually use these tools and considering the potential downstream risks to privilege that GenAI tools present.

  • Consider whether you have implemented enterprise GenAI tools, including applications that have incorporated AI functionality, and whether employees are permitted to use publicly available GenAI tools for business purposes.

    • Many companies are implementing enterprise versions of common GenAI tools like Claude which, when configured correctly, are confidential to the company and do not permit third parties like Anthropic to monitor the content of prompts and outputs generated by these enterprise tools. The use of enterprise GenAI tools by companies and external law firms helps to safeguard the confidentiality of the prompt and output and helps to minimize the risk that parties have waived privilege because the information is not confidential as between the client and their counsel. Locations where the legal team stores privileged information or matter-specific materials should not be included in the AI models that are available to all employees.
    • Even when your company is using an enterprise version of these tools, your employee policies and AI governance program should provide employees with guidance regarding the use of AI tools, including what information can be included in the AI tool, requiring employees to review the outputs and describing how AI outputs can be used and stored.
    • Companies should prohibit the use of public AI tools by employees for work purposes that involve confidential information, including privileged information, to minimize the risk of privilege waiver (or leakage of company confidential information).
  • Consider whether your organization permits meeting transcripts and summaries created by AI tools.

    • If you allow transcripts or recordings, consider prohibiting recordings, transcripts and AI summaries of meetings with any privileged content given the sensitivity of the material, and the potential that these outputs and prompts may need to be collected and logged in discovery.
    • For non-scripted meetings, consider setting a reasonable retention period for the recordings and transcripts (with an exception for individuals subject to legal holds). Default retention settings are often indefinite, so if no retention period is applied, large volumes of recordings and transcripts may be potentially available for preservation and collection.
    • For meeting summaries, encourage employees to review the summaries to ensure they are accurate.
  • GenAI can be a tool for legal practitioners conducting pre-suit and early case assessment, but it should be specifically and expressly directed and guided by counsel.

    • It is always more challenging to demonstrate that materials generated by non-lawyers are privileged and were in fact created with the purpose of seeking or obtaining legal advice than materials generated by lawyers.
    • Consider implementing policies and procedures that direct employees not to use GenAI tools for purposes related to legal matters – with specific guidance to legal hold custodians – absent the direction of counsel. This guidance would include inputting legal advice, and also seeking to generate information related to legal matters.
    • In turn, legal practitioners should thoughtfully consider how they intend to use GenAI tools, or direct others to use GenAI tools, to support legal investigations. This includes considering whether the tool is confidential, providing specific direction on whether GenAI tools may be used to prepare materials and the risk that certain prompts and outputs may be discoverable.

Our team has significant experience dealing with GenAI information, from both a governance perspective on the proper management of GenAI from its creation to storage and disposition, and a litigation and e-Discovery perspective, including strategies for preservation, collection, negotiation and disclosure, including defending privilege claims. For any questions regarding issues relating to those matters, please contact us.

Footnotes

  1. 1

    Memorandum, United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y. Feb. 17, 2026), Dkt. No. 27.

  2. 2

    In re OpenAI, Inc., Copyright Infringement Litig., No. 25-MD-3143 (SHS) (OTW), 2026 WL 21676, at *3 (S.D.N.Y. Jan. 5, 2026).

  3. 3

    Concord Music Grp., Inc. v. Anthropic PBC, No. 24-CV-03811-EKL (SVK), 2025 WL 1482734, at *2-3 (N.D. Cal. May 23, 2025).

Contact