On February 17, 2026, a United States District Judge for the Southern District of New York issued a landmark decision regarding the use of publicly-available generative AI tools in litigation. This opinion, while not binding in Texas, may be highly persuasive and confirms concerns that many lawyers have raised given the rise in accessibility of such AI platforms.
In the case of U.S. v. Heppner, the government seized defendant’s computers when he was arrested. They later learned that after the defendant was aware of the investigation, but before his arrest, the defendant had submitted various queries and prompts to Claude, Anthropic’s publicly-available generative AI platform. Defense counsel asserted that the documents generated by Claude in response to those prompts were privileged attorney-client communications and/or protected attorney work product that that government were not permitted to access, view, or utilize in the prosecution. The government disagreed, and the Court ultimately sided with the government.
The Court rejected the claim of attorney-client privilege, finding that the documents were not communication between the defendant and his attorney. The Court also referred to Claude’s privacy policy, which expressly states that Anthropic collects and stores data on inputs and outputs, which it can use to train Claude. Anthropic also reserves the right to disclose that data to a variety of third parties. Therefore, the defendant’s communications with Claude were not confidential. Additionally, the Court determined that the defendant was not communicating with Claude for the purpose of receiving legal advice. The Court was particularly persuaded by Claude’s disclaimer that it does not provide legal advice, noting that when the government asked Claude whether it could give legal advice, the program replied that it was not a lawyer, could not give legal advice, and it recommended the user confer with “a qualified attorney.” Relying on long-standing interpretation of attorney-client privilege, the Court concluded that the defendant could not transform these non-privileged communications into privileged communications by later sending them to his attorneys. Finally, in response to defendant’s argument that the queries posed to Claude contained information that was communicated to the defendant from his attorneys and therefore privileged, the Court ruled that the defendant waived the privilege by disclosing that content to Claude.
The work product doctrine is somewhat related to the attorney-client privilege, but it is designed to protect the mental processes and impressions of the attorney. The Court ruled that the documents generated by Claude were not protected from disclosure under this theory because they were not “prepared by or at the behest of counsel” and they did not reflect the legal strategy of counsel.
Since the explosion of the use of generative AI in the last few years, many attorneys have been concerned about its implication in the representation of clients. Many attorneys have been warning that the use of public platforms or programs was risky for exactly the reasons outlined in the Heppner case—clients may be unwittingly waiving attorney-client privilege and risking that their litigation strategy will be disclosed to opposing parties. While these tools seem powerful, they are not well suited for providing accurate legal information and the adverse consequences of feeding privileged communications into these public platforms are significant. If you have questions, please contact your L&B attorney.
