By Hunter Steele
Attorney-client privilege and work-product protection are distinct legal doctrines that have both been shaped by case law and rule generation over time. This shaping continues as both doctrines are applied to modern technology such as artificial intelligence.
Attorney-client privilege protects confidential communications between client and lawyer made to obtain or provide legal advice.[1] Attorney-client privilege protects confidential communications made to obtain legal advice but does not make preexisting, otherwise producible documents privileged merely because they are delivered to an attorney.[2] Further, under Kovel doctrine, privilege may extend to a nonlawyer agent whose assistance is necessary to facilitate legal advice.[3] These rulings are incorporated into Federal Rule of Evidence 501, which applies federal common-law principles guided by “reason and experience.”[4]
Work product serves a different function. Hickman v. Taylor recognized a zone of privacy for counsel’s litigation preparation.[5] Further, Federal Rule of Civil Procedure 26(b)(3)(A) codifies protection for documents and tangible things prepared by or for a party or its representative in anticipation of litigation, and Rule 26(b)(3)(B) protects an attorney’s mental impressions, conclusions, opinions, and legal theories. Work product therefore protects litigation preparation, not every confidential client communication. For generative artificial intelligence (AI), the result depends on who directed the use, why the material was created, and whether disclosure exposed it to an adversary.
In United States v. Heppner, the court considered 31 documents that a criminal defendant created with AI model Claude while preparing possible defenses to a federal securities and wire-fraud investigation.[6] Heppner used Claude without counsel’s direction, then gave the materials to his lawyers. The court ordered production and ultimately rejected privilege because Claude was not an attorney or licensed professional, the exchanges were not communications with counsel for legal advice, and the public platform’s terms defeated any reasonable expectation of confidentiality.[7] Sharing the outputs with counsel did not make them privileged.[8]
Work-product protection for generative AI use failed separately in Heppner. Even assuming anticipation of litigation, the documents at issue were not prepared by or at counsel’s behest and did not reveal defense counsel’s strategy.[9] Although Kovel permits protection for a qualifying agent assisting counsel, Heppner emphasized that counsel had not directed Claude’s use.[10]
In Warner v. Gilbarco, a pro se employment plaintiff used ChatGPT to answer legal questions and draft filings.[11] Defendants sought the prompts and responses. Magistrate Judge Anthony P. Patti treated the materials as opinion work product because they reflected Warner’s mental impressions and litigation strategy. The court viewed ChatGPT as a tool, not a person, and held that use alone did not waive work-product protection.[12] Warner follows Hickman and Rule 26(b)(3), while extending their protective purpose to a self-represented litigant. Protecting a pro se party’s strategic preparation preserves the adversarial system’s “zone of privacy.”[13] Unlike Heppner, it did not ask whether ChatGPT supplied legal advice or formed an attorney-client relationship.
Viewed together, the decisions are not irreconcilable. Prior law asks different questions: was there a confidential client-lawyer communication made to obtain legal advice? Were materials prepared by or for a party or representative because of anticipated litigation, and do they embody protected strategy or mental impressions against an adversary?
Heppner answered no to both questions: the defendant consulted a public AI platform independently, and his materials did not reflect counsel’s direction or strategy. Warner answered yes to the work-product question: the pro se plaintiff used ChatGPT for litigation preparation, without disclosure to the adversary. The contrast turns on the doctrine invoked, the user’s role, the material’s purpose, counsel’s involvement, and the risk of adversarial access. Generative AI changes the medium; it does not erase these established requirements.
Hunter Steele is a 2L at Vanderbilt Law School. He plans on working with emerging technology companies after law school.
[1] Upjohn Co. v. United States, 449 U.S. 383, 389–90 (1981).
[2] Fisher v. United States, 425 U.S. 391, 403–04 (1976).
[3] United States v. Kovel, 296 F.2d 918, 921–22 (2d Cir. 1961).
[4] Fed. R. Evid. 501.
[5] Hickman v. Taylor, 329 U.S. 495, 510–11 (1947).
[6] United States v. Heppner, No. 25 Cr. 503 (JSR), 2026 WL 436479, at *1–4 (S.D.N.Y. Feb. 17, 2026).
[7] Id.
[8] Id. at 2–3; see Fisher, 425 U.S. at 403–04 (1976).
[9] Heppner, 2026 WL 436479, at *3.
[10] Id.
[11] Warner v. Gilbarco, Inc., No. 2:24-cv-12333, 2026 WL 373043, at *4, *11–13 (E.D. Mich. Feb. 10, 2026).
[12] Id. at *11–13.
[13] Hickman, 329 U.S. at 510–11.