Before privilege or production issues arise, the more basic inquiry is whether the information was ever confidential in the first place. This question is particularly significant in e-discovery, where electronically stored information generated by AI systems may later become discoverable.
In the March 19, 2026 edition of The Legal Intelligencer, Kelly Lavelle writes, “AI Systems and the Question of Confidentiality.”Much of the attention surrounding generative AI in litigation has focused on hallucinated authorities and AI-generated filings. But a different litigation risk arises when lawyers, or even clients themselves, enter client information into an AI platform. Before privilege or production issues arise, the more basic inquiry is whether the information was ever confidential in the first place. This question is particularly significant in e-discovery, where electronically stored information generated by AI systems may later become discoverable.
The U.S. District Court for the Southern District of New York’s decision in United States v. Heppner brings that issue into focus. Although the case has been described as a privilege ruling, the court’s analysis turned on a more basic issue: whether the information was confidential when it was entered into the AI platform. If client information is submitted to a system under terms that allow the provider to access, retain, or disclose it, the protections that depend on confidentiality, including attorney-client privilege and work product, may never attach.
In Heppner, a corporate executive under criminal investigation used the publicly available generative AI platform “Claude” to generate written analyses of potential defenses after receiving a grand jury subpoena. He later shared those documents with counsel. Following his arrest, FBI agents executed a search warrant at Bradley Heppner’s home and seized numerous documents and electronic devices. Heppner’s counsel later informed the government that among the seized materials were approximately thirty-one documents memorializing Heppner’s communications with the AI platform.
Through his counsel, Heppner asserted privilege over these documents, arguing that the information entered into the AI platform was learned, that he created the documents to facilitate discussions with counsel and obtain legal advice, and that he subsequently shared the AI outputs with his attorneys. His counsel acknowledged, however, that they did not direct Heppner to conduct the AI searches. The court rejected those claims. It held that the documents were not communications with an attorney, were not prepared by or at the direction of counsel, and, critically, were not confidential in light of the platform’s privacy policy.
The court’s analysis centered on the conditions under which the information was entered into the system. The provider’s terms stated that user inputs and outputs were collected and retained and that the company reserved the right to disclose data to third parties, including governmental authorities. Under those conditions, the court concluded that Heppner lacked a reasonable expectation of confidentiality at the moment of disclosure.
The court also rejected Heppner’s contention that he communicated with Claude for the “express purpose of talking to counsel.” In doing so, the court looked to the platform’s own representations. Claude expressly disclaimed providing legal advice. When asked whether it could provide such advice, the system responded that it was not a lawyer, could not offer formal legal recommendations, and advised users to consult a qualified attorney.
The court acknowledged that the implications of artificial intelligence for the law are only beginning to be explored. The court emphasized that AI’s novelty does not place it outside established legal principles, including those governing attorney-client privilege and the work-product doctrine. AI does not change the basic rules.
The opinion highlights a familiar but often underemphasized principle that privilege attaches only when communications are intended to be confidential and were, in fact, kept confidential. In the AI context, if a platform’s terms allow provider use beyond delivering the requested service, such as model training, analytics, affiliate sharing, or extended retention, the user may be acting inconsistently with the intent to preserve confidentiality at the moment of disclosure. The issue is not a subsequent waiver. It is that the platform itself may defeat confidentiality at inception.
This reasoning extends beyond generative AI. Courts have consistently held that employees may waive attorney-client privilege when using employer email systems to communicate with counsel, with the analysis turning on whether the employee had an objectively reasonable expectation of confidentiality. In those cases, the inquiry focuses on whether the intent to communicate in confidence was objectively reasonable under the circumstances. Heppner suggests that AI platforms may be subject to similar scrutiny. The terms governing data retention and disclosure may determine whether any protection exists at all.
The court also reaffirmed that privilege remains grounded in professional accountability. Recognized privileges depend on a relationship with a licensed professional who owes fiduciary duties and is subject to discipline. Generative AI does not occupy that role. Even if an output resembles legal advice, it is not a communication with counsel.
However, the court suggested that if counsel had directed the use of the AI tool within a structured framework, the analysis might have been different. That distinction highlights the importance of control. When AI use is integrated into a counsel-directed process under defined terms, protection arguments may be stronger. When it is independent and unsupervised, confidentiality may fail at the start.
Platform terms of service are therefore central. Many providers claim licenses to user inputs and outputs broader than a confidentiality-preserving relationship would tolerate, sometimes including rights to use content for service improvement, analytics or model training. However, this concern does not apply in the same way to all AI tools. Paid legal research platforms or firm-licensed systems are often governed by professional services agreements that include confidentiality obligations and defined data controls. In those contexts, the vendor functions more like a traditional litigation support provider than a public AI service. Even there, however, the analysis may turn on the contract. Courts will look at the terms to determine what they permit and how the provider handles user content.
These issues directly affect attorney-client privilege. Privilege protects confidential communications made for the purpose of obtaining or providing legal advice. The problem is not simply waiver. It is that the basic requirements for privilege may never have been satisfied. The work-product doctrine raises a similar issue. Work product protects materials prepared in anticipation of litigation. If those materials are created on a public nonconfidential platform, it weakens the argument that they reflect protected legal strategy shielded from disclosure.
Heppner reinforces the basic rule that privilege depends on confidentiality at the time the communication is made. It does not arise simply because a document is later shared with counsel. In the AI context, that means platform terms, data handling practices, and professional controls must align with confidentiality requirements before client information is entered. The focus should be less on whether AI interactions are discoverable and more on whether confidentiality ever attached. As Heppner shows, confidentiality is not assumed simply because a document resembles legal analysis. It depends on the conditions under which the information was created and maintained.
Kelly A. Lavelle is Senior Counsel at Kang Haggerty. She focuses on e-discovery and information management, from preservation and collection to review and production of large volumes of electronically stored information. Contact her at klavelle@kanghaggerty.com.
Reprinted with permission from the March 19, 2026 edition of “The Legal Intelligencer” © 2026 ALM Global, LLC. All rights reserved. Further duplication without permission is prohibited. Request academic re-use from www.copyright.com. All other uses, submit a request to asset-and-logo-licensing@alm.com. For more information visit Asset & Logo Licensing.