The use of AI and similar evolving technologies by lawyers in the United States implicates significant and complex privilege issues. As noted above, the attorney-client privilege protects confidential communications between a client and their attorney made for the purpose of securing legal advice. While various US states may have nuanced differences in their respective approach to the attorney-client privilege may vary in its nuances across US states, all recognize that the attorney-client privilege may be waived by disclosure to third parties. Applied to AI, in general, entering the contents of privileged attorney-client communications into public generative AI models could be deemed a failure to keep the communications confidential by disclosing them to a third party, resulting in a waiver of the privilege protection. Public AI models often "learn" from user inputs. Information contained in a user’s input/prompt gets incorporated into the AI model’s potentially public records and can even appear in another user’s output/response. These technical features and details may result in an inadvertent third-party disclosure which can constitute a waiver of the privilege. Even private, vendor-hosted AI models can pose risks if they are programmed to learn from user inputs, as courts might decline to protect communications whose contents could be disclosed, for example, to other firm clients or the vendor’s AI trainer.

Lawyers can, however, mitigate the risk of privilege waiver by ensuring AI models maintain strict confidentiality (i.e., by building those models in a closed, internal environment behind a firewall and implementing other appropriate controls such as not programming the model to learn from user inputs). That said, the law in this area continues to develop, and lawyers should proceed cautiously.

The other privilege protection in the United States, the work product doctrine, provides protection to written or oral information prepared in preparation for litigation or for trial. The laws governing the work product doctrine also vary among the states, but in general the doctrine’s protections would appear to apply to content, e.g., memos, prepared by attorneys using generative AI in the same fashion as content created without the use of AI, provided the AI model has the same strict confidentiality safeguards mentioned above. As illustrated in a recent case, the work product doctrine may protect certain inputs, prompts and outputs into AI tools and models.¹ Lawyers should take limited comfort from this early decision in light of uncertainty in the application of the work product doctrine, litigating parties’ ability to overcome the doctrine’s protections for fact work product, the many open legal questions that remain unanswered, and the continued evolution of AI technology.. 

The use of generative AI tools similarly raises ethical concerns for lawyers who are obligated to maintain client confidentiality. In the United States, the American Bar Association (ABA) provides guidance to attorneys about the attorney-client privilege. In particular, on July 29, 2024, the ABA released its Formal Opinion 512 on the use of generative AI in legal practice, emphasizing the importance of maintaining confidentiality and obtaining informed consent when using these tools. ABA Model Rule 1.6 (Confidentiality) addresses key points regarding the attorney-client privilege. It mandates vigilance in protecting client information, including when using AI tools. Rule 1.4 (Communication) may require client consultation about AI use in their matters, particularly when confidentiality concerns exist. As the US federal government, US state governments and various bar associations continue to grapple with the ethical and practical implications of rapid advances in AI technology on legal privileges, readers are urged to closely monitor these developments.

[1] See Tremblay v. Openai, Inc., No. 23-cv-03223-AMO, 2024 U.S. Dist. LEXIS 141362 (N.D. Cal. Aug. 8, 2024).