Quarter 1, 2026: W(AI)ving The Privilege: AI Discoverability
By: Attorney James Rayment and Shareholder Spencer C. Pittman
Whether inputs and prompts shared with generative AI tools fall within the scope of evidentiary privileges, such as the attorney-client privilege or work product, remains an unsettled and increasingly consequential question. This issue has resulted in conflicting rulings across the country and is causing courts and practitioners to navigate uncharted territories.
The attorney-client privilege is a foundational legal principle that protects certain communications between a client and an attorney. An exception to the attorney-client privilege is when an otherwise privileged communication is disclosed to or in the presence of a third party, such as friends or family. An adjacent principle to the attorney-client privilege is the “work product doctrine,” which protects materials prepared in anticipation of litigation. Popular generative AI tools, such as ChatGPT, Grok, and Claude, are generally owned and operated by technology companies. If you share information with an AI tool, you may be directly sharing it with those technology companies, a third party, to train future AI-models. This leads to the question of whether AI inputs waive privileges. Courts are now beginning to address this question, but the answers are not unanimous.
In U.S. v. Heppner (25 Cr. 503, S.D.N.Y.) the defendant (a non-lawyer) was indicted by a grand jury. Police executed a search warrant and obtained, among other items, 31 documents that included the defendant’s AI-inputs into a free-version of Claude, a generative AI tool. The defendant’s use of Claude, before his retention of counsel, was for independent research and strategizing for his criminal defense along with exploring what possible criminal charges could be forthcoming. The defendant’s counsel argued the AI inputs were work product and not discoverable. The government disagreed with the defendant’s argument and sought the production of the AI inputs. Although the defendant subsequently had retained counsel and shared the AI prompts, the Court found that, in this case, the communications between the defendant and the publicly available AI tool were not protected by either the attorney-client privilege or the work product doctrine. The Court ruled that neither the attorney-client privilege nor the work product doctrine applied because the inputs were put into a free version of Claude, which (unlike certain paid or enterprise versions) “learned” or trained itself from user inputs. The Court also found the work product doctrine did not apply because the inputs were not used for the purpose of obtaining legal advice, were not prepared by or at the behest of his attorney, and did not reflect the defense counsel’s strategy.
However, the same day as the Heppner ruling (Feb. 10, 2026), the E.D. of Michigan reached the opposite conclusion in Warner v. Gilbarco, Inc. (Case No. 2:24-cv-12333), an employment-related civil action. The plaintiff in Warner was a pro se litigant representing herself . The defendants were denied the plaintiff’s AI inputs by the Court because Warner used the AI-inputs in “anticipation of litigation” (i.e., work product) and despite using presumably the free-version of ChatGPT, the Court found the waiver of work product must be to the adversary or likely to get into an adversary’s hands.
Courts in Oklahoma and within the 10th Circuit have not offered guidance on whether those Courts would take the Heppner or the Warner approach and find AI-tools are or are not “third parties” for purposes of work product, but one key takeaway is certain- expectations of privacy in AI inputs should be questioned at all times and will depend on a variety of factors, such as (i) counsel’s involvement, (ii) the GenAI’s privacy policy/terms, and (iii) the jurisdiction of the Court.
Instead of relying on AI to advise you or your business on your legal rights, contact the litigation team at Winters & King at 918-494-6868