As everyone uses AI more and more for everything, AI use records have become the latest electronically stored information (ESI) discovery darling. Consider the corporate CEO deponent confronted at deposition with his Claude prompt, “What does a CEO do?” – real case and yes, really discoverable. Or, consider these real-world examples straight from my arbitrations:
A manager asking ChatGPT, “How do I terminate a pregnant employee without getting sued?”
An HR Director inputting – without legal’s direction -confidential details about his internal investigation into an employee’s sexual harassment complaint to evaluate the company’s exposure under Title VII.
After receiving an EEOC charge but before contacting counsel, a senior executive questions generative AI about the claims and the company’s concerns and then asks AI to evaluate the strengths and weaknesses of defending the claim.
Without a lawyer in the loop, every AI use case could create potentially discoverable evidence. But even with a lawyer in the loop, not everything will be privileged. Even with lawyer involvement, privileged advice could potentially be waived if run through AI systems that do not have confidentiality protections. So far, discovery of AI evidence is increasing but not yet commonplace. Both sides of the bar, however, are actively seeking AI evidence in discovery – especially in arbitrations, which enjoy a greater privacy cloak. Two recent cases suggest how courts will approach attorney-client privilege and attorney work product protections in the discovery of AI-generated materials and underscore the risk of AI generated materials potentially becoming “bad evidence.”
United States v. Heppner, 828 F. Supp. 3d 484 (S.D.N.Y. 2026)
In Heppner, a criminal defendant used a generative AI tool, Claude, to prepare reports analyzing potential charges and outlining possible defense strategies. He did this independently, without the direction of counsel, and later shared the AI-generated materials with his attorney. After learning of Heppner’s AI use, the government requested those materials in discovery. Heppner’s attorney argued that the attorney-client privilege and work product doctrine protections. The court rejected both arguments.
Attorney-Client Privilege: The court found that the materials were not privileged. First, the court found that the AI platform was not an attorney. Second, the court found that the materials could not be privileged because Heppner used a non-confidential AI tool. The court highlighted the AI provider’s terms of service, which expressly disclaimed the provision of legal advice and reserved the right to use and disclose user data. Accordingly, the court found Heppner lacked a reasonable expectation of confidentiality for the communications to be privileged. Third, the court found that Heppner did not use the AI tools for purposes of obtaining legal advice since he did not do so at the direction of counsel. The court also rejected the idea that sharing the materials with counsel after the fact could magically transform them into privileged communications.
Attorney Work Product: The court emphasized that the attorney work product doctrine exists to protect the mental impressions and strategic thinking of attorneys. Although work product can extend to materials prepared by non-lawyers, the protection depends on whether the material reflects counsel’s strategy prepared at counsel’s direction. Here, the defendant acted on his own initiative. The AI-generated reports may have influenced later discussions with counsel, but they did not reflect counsel’s thought processes at the time they were created. That distinction proved decisive to the court’s conclusions.
Warner v. Gilbarco, 820 F. Supp. 3d 629 (E.D. Mich. 2026).
In Warner v. Gilbarco, a pro se employment plaintiff used ChatGPT to assist with drafting legal documents and for litigation strategy. The defendant sought to compel production of all materials related to her use of AI in connection with the lawsuit.
The court found the AI-generated materials protected by the attorney work product privilege because they reflected plaintiff’s mental impressions, drafting process, and litigation strategy. As a pro se litigant, plaintiff essentially acted as her own lawyer. As such, the court found that those materials constituted work product. The court also characterized generative AI platforms as “tools, not persons,” and distinguished work product waiver from attorney-client privilege waiver. The court found no waiver of the work product doctrine because plaintiff had not disclosed her mental impressions to an adversary or in a manner likely to place the materials in an adversary’s hands. Notably, the court did not address whether plaintiff had used a non-confidential version of ChatGPT that the Heppner court had found could result in a waiver.
Lessons Learned?
Courts are treating AI-generated materials as ordinary evidence subject to ordinary discovery rules. Courts are applying familiar rules for attorney-client privilege, work product doctrine, and waiver to the new AI technological realities. Consequently, using AI without lawyer involvement and not at counsel’s direction may make the attorney-client privilege and/or work product doctrine inapplicable. Sending those AI materials later to counsel will not automatically cloak them with privilege. Creating AI materials on public versions of generative AI models will likely waive any attorney-client privilege and/or work product protection that apply.
As an arbitrator, expect the new frontier to arrive at your door with parties seeking discovery of:
Generative AI prompts and outputs related to the claims;
AI-generated drafts;
Collaborative platform outputs;
Cloud-based project management systems;
Meeting recordings, transcripts, and summaries created by AI assistants;
All of the above on personal devices used for work or in connection with a claim; and
Deposition questions on AI tools, use, and outputs.
Also expect to see:
More pro se claimants using AI to file claims;
Increasing spoliation claims requesting sanctions; and
A need for a deeper understanding of closed system protections.
This article is the third in a series. Read Part 1 here and Part 2 here.
Janice L. Sperow is a member of the AAA's AI Ambassador program, which brings together experienced AAA arbitrators and mediators to examine emerging issues at the intersection of AI and dispute resolution.