Chats may not be privileged

United States lawyers are telling clients to stop treating chatbot conversations like private talks with counsel after a federal judge ruled those exchanges can be handed over in court. (money.usnews.com) The warning follows a February 17, 2026 decision by Judge Jed Rakoff of the Southern District of New York in *United States v. Bradley Heppner*. Rakoff held that a criminal defendant’s written exchanges with a publicly available generative artificial intelligence platform were not protected by attorney-client privilege or the work product doctrine. (perkinscoie.com) The case involved about 31 documents created after Bradley Heppner learned he was the target of a criminal investigation and, on his own initiative, used Anthropic’s Claude to sketch out defense arguments. Federal agents later seized the materials, and the court said sharing them with a lawyer afterward did not make them privileged. (harvardlawreview.org) Attorney-client privilege is the rule that usually keeps confidential legal communications between a client and a lawyer out of court. Rakoff said the chatbot was not a lawyer, and the chats were not confidential because the service’s policy said it collected inputs and outputs, used data to train the model, and could disclose data to third parties, including government authorities. (harvardlawreview.org; thompsonhine.com) Law firms are now pushing that warning into client advice and engagement letters. Reuters reported on April 15 that some firms are explicitly telling clients not to put legally delicate or commercially sensitive facts into external chatbot tools. (money.usnews.com) The ruling does not say every use of artificial intelligence destroys privilege. Several firms said Heppner turned on specific facts: he used a public tool, without direction from counsel, and the platform disclaimed privacy, leaving harder questions for lawyer-supervised use and locked-down enterprise systems. (dlapiper.com; goodwinlaw.com) That distinction matters because consumer and business products are treated differently by many providers. OpenAI says content from its consumer services may be used to improve models, while data processed for business offerings is governed by separate customer agreements. (openai.com) Anthropic has also drawn a line between consumer and commercial use in its privacy materials. The company said in a December 17, 2024 update that it processes personal data on behalf of commercial users as a data processor, while separate training notices explain how personal data may be used in model training. (privacy.claude.com; privacy.claude.com) For companies, the practical shift is toward tighter rules on who can use outside chatbots and what can be pasted into them. For clients, the safer assumption after Heppner is that a chatbot log can look less like a confidential legal memo and more like a document the other side may try to obtain. (crowell.com; foxrothschild.com)