Key Takeaways
Courts are increasingly confronting requests for party communications with AI platforms and are applying traditional privilege and work-product principles to resolve them. Key facts likely to drive the analysis include the nature of the relationship between the user and the AI platform, whether counsel directed the use of the AI tool, whether the platform’s terms of service preserve the confidentiality of user inputs, and whether the prompts or outputs reflect counsel’s mental impressions or litigation strategy.
U.S. District Judge Jed S. Rakoff held in a criminal fraud case that a financial services executive could not claim privilege over documents memorializing his communications with the AI tool, Claude.
The defendant asserted privilege and work-product protections over documents that the FBI seized in its investigation. The defendant argued that he input information learned from his attorney into Claude, which generated responsive output, that he created the documents for the purpose of speaking with counsel to obtain legal advice, and that he then shared those documents with counsel. It was undisputed that defendant’s counsel did not direct the defendant to use AI.
Judge Rakoff held that the AI documents were outside the attorney-client privilege: according to the decision they were not attorney-client communications; they lacked confidentiality because the AI platform was a third party, and Claude users consent to Anthropic’s privacy policy that allows disclosure of inputs and outputs to additional third parties, including governmental authorities; and they lacked the purpose of obtaining legal advice where the defendant’s counsel did not direct or suggest that defendant use AI. Judge Rakoff likewise held that the AI documents did not merit work-product protection because, even assuming they were prepared in anticipation of litigation, they were not prepared by or at the behest of counsel, nor did they reflect defense counsel’s strategy.
Although Judge Rakoff described the ruling as addressing a question of first impression, the court did not purport to create new law. Rather, the court stated that it applied “longstanding legal principles . . . governing the attorney-client privilege and the work product doctrine,” concluding that the defendant’s use of AI failed to satisfy either doctrine. Other cases have also treated AI prompts and outputs as a form of electronically stored information subject to ordinary discovery principles. See, e.g., Concord Music Grp., Inc. v. Anthropic PBC, 2025 WL 1482734, at *2 (N.D. Cal. May 23, 2025).
The ruling leaves open the possibility that AI-generated prompts and outputs do qualify for privilege or work-product protection under different circumstances. Indeed, courts have concluded that prompts and AI settings can constitute attorney work product. Concord, 2025 WL 1482734, at *2. For example, in contrast to Judge Rakoff’s recent decision, Warner v. Gilbarco, Inc. reached the opposite result, denying a motion to compel production of a pro se party’s use of third-party AI tools in connection with litigation, finding the information not discoverable, not proportional to the needs of the case, and protected as work product. 2026 WL 373043, at *3-5 (E.D. Mich. Feb. 10, 2026). The court further rejected the argument that the pro se litigant waived work-product protection by using AI, reasoning that work-product waiver requires disclosure to an adversary or in a manner likely to reach an adversary, and that “generative AI programs are tools, not persons.” Id. at *4.
