AI-Generated Content Not Protected by Legal Privilege
A U.S. federal court in the Southern District of New York has ruled that communications with an AI model are not protected by attorney-client privilege. The *United States v. Heppner* decision found that inputting sensitive information into a third-party AI waives confidentiality. In response to such risks, California's Senate passed SB 574, a bill requiring lawyers to obtain client consent before using AI for substantive work.
- In the *United States v. Heppner* case, the defendant, Bradley Heppner, used the free, public version of Anthropic's Claude AI to generate and outline potential defense strategies after he was aware of a federal investigation but before his indictment on fraud charges. - The court's decision, issued by Judge Jed S. Rakoff, hinged on the fact that Heppner used a consumer-grade AI tool on his own initiative, not at the direction of his lawyers. This independent use was a key factor in ruling that the communications were not protected by the work-product doctrine. - A critical element in waiving the attorney-client privilege was the AI's terms of service, which permitted data collection, retention, and use for model training, thereby negating any reasonable expectation of confidentiality. By inputting information that his attorneys had previously shared with him, Heppner was deemed to have waived the privilege over those original communications as well. - The court rejected the argument that the AI served as a privileged intermediary, like a translator or accountant, under the *Kovel* doctrine, because the AI was not necessary for the lawyers to understand their client and was not engaged at the counsel's direction. - The ruling does not prevent attorneys from using AI altogether; many legal commentators suggest that using secure, private, or enterprise-level AI platforms under the direct supervision of an attorney could still preserve privilege protections. - California's SB 574, which passed the Senate 39-0, would codify existing bar guidance into law, requiring lawyers to verify the accuracy of AI-generated material, correct "hallucinated" or biased output, and prevent the entry of confidential client information into public AI systems. - The California bill also seeks to regulate AI in alternative dispute resolution, prohibiting arbitrators from delegating any part of their decision-making process to a generative AI tool.
