A recent buzz going around in legal circles is whether conversations with AI on a matter that finally ends up in court are privileged or not—that is, whether they are discoverable by the opposing party. Two recent cases shed some light on this question.
The Heppner Case: When AI chats lost Privilege
On February 17, 2026, United States District Court Judge Jed S. Rakoff of the Southern District of New York issued a written memorandum in United States v. Heppner, No. 25 Cr. 503 (S.D.N.Y.). He declared that unsupervised client use of a non-enterprise Generative AI platform defeated attorney-client and attorney-work-product privilege, ordering the disclosure of defendant Bradley Heppner’s musings with his chatbot to the government. Heppner had been arrested on wire fraud charges. When agents searched his residence, they seized electronic devices containing chats with Claude initiated after he received a grand jury subpoena. Heppner argued that he chatted with Claude based on information he learned from his lawyer to find talking points for their meetings, and because he eventually shared those conversations with his lawyer, they should be privileged.
The Court disagreed and ordered the chats turned over. Judge Rakoff noted that these memorialized documents were not communications between Heppner and his counsel. Furthermore, an interaction with Claude is not a conversation with a “trusted partner.” Most importantly, the Court held the conversations were not confidential because Heppner used the public Claude platform. Anthropic’s privacy policy states that prompts are retained, used for model training, and can be disclosed to third parties, including the government.
The Warner Case: A Shield for Pro-Se Litigants
In the second case, United States Magistrate Judge Anthony P. Patti of the Eastern District of Michigan issued a ruling in Warner v. Gilbarco, Inc. (Feb. 10, 2026). He denied a defendant’s request for disclosure of all documents related to the plaintiff’s use of ChatGPT. The Warner case is an employment dispute where the plaintiff represented herself. As a pro-se litigant, she used ChatGPT to assist with her case. When the opposing side sought access to these conversations, the Court refused, stating that materials prepared in anticipation of litigation are protected. The Court observed that while Warner’s use of ChatGPT might have compromised attorney-client privilege, it did not destroy the work-product privilege because she was representing herself, the materials were prepared for the litigation, and she did not share them with an adversary. The Court reminded everyone that ChatGPT and similar platforms are just tools, not persons.
Weakest Link: The Attorney-Client Privilege
While some commentators suggest these decisions are inconsistent, I see common themes that will likely resonate in future cases. Both decisions show that attorney-client privilege is the weakest link when using Generative AI for litigation, especially when using “freemium” tools that lack privacy protections and use data for training. However, this is not a deal-breaker. Both courts observed that the attorney-work-product privilege may act as a shield even when using unprotected tools. This is particularly true for pro-se litigants handling their own cases.
As Generative AI becomes ubiquitous, more people will naturally use tools like ChatGPT, Claude, or Copilot to understand their legal situations. Law firms often advise clients to simply not use these platforms or upload materials, but that often ignores the ground reality of how people work today. Such rigid advice may hold water during active litigation, but it won’t stop the flood of usage in pre-litigation scenarios and might prevent litigants from becoming informed customers.
Practical Advice: How to Protect Yourself from Disclosure
So, how should litigants and businesses navigate this?
First, be aware of the privacy policies and data sharing terms of your AI tools. Understand that freemium tools generally offer no confidentiality. Use these tools to learn about the generality of a situation rather than your specific facts. Be cognizant that chatting from your office or home may not provide protection from later disclosure unless you are representing yourself.
If you decide to use a tool for anticipated litigation, use a professional subscription. In your prompts, embed the intent of the session by including statements like “prepared in anticipation of litigation to discuss with attorney” or “prepared at the request of counsel.” Regardless of your subscription level, never upload highly confidential documents. If a court eventually orders disclosure, those documents would go straight to your adversary. Finally, if you are a medium-sized business or larger, implement an AI governance framework and ensure your managers are aware of it.
At ANRLAWLA we take privacy seriously, we are ever vigilant, and help our clients stay compliant. If you need to do a privacy audit of your business, website or software as a service offering please reach out to us at [email protected].
Leave a Reply