Choosing a provider
Court: defence documents made with Claude are not privileged
In February 2026 a federal court in New York ruled that documents a defendant created with Claude on his own initiative were not protected by legal privilege. Anthropic's privacy policy played a part in the decision.
In 2025, after receiving a grand jury subpoena and learning he was the target of an investigation, a businessman later charged with securities fraud used Anthropic's AI assistant Claude to prepare reports on his possible defence. His lawyers had not asked him to. They said some 31 such documents were among material the FBI seized at his home when he was arrested in November 2025. The defence claimed privilege: he had entered information from his lawyers, had prepared the documents to discuss with them, and had later shared them.
The US District Court for the Southern District of New York rejected this from the bench on 10 February 2026 and gave written reasons on 17 February. Claude is not a lawyer, the court said, and that alone defeats attorney-client privilege. Nor were the exchanges confidential, not only because they were with a third party but because Anthropic's privacy policy says it collects inputs and outputs, uses them to train Claude, and may disclose them to third parties, including government authorities. Any privileged information typed in lost its protection by being shared, and handing the documents to counsel later did not make them privileged. Work-product protection failed too, since the documents were not prepared at counsel's request.
The court noted that had counsel directed the use of Claude, the tool might arguably have been treated like a lawyer's agent. It saw the question as one of first impression nationwide. This is a single first-instance ruling; lawyers at Debevoise & Plimpton argue that confidential enterprise AI tools that do not train on inputs should be treated differently.
What can we learn from it?
The ruling does not say that AI use rules out confidentiality; it says long-standing rules apply to it too, and the provider's terms matter: where they say inputs are not confidential, the court did not treat them as such. The court also indicated that a tool used at a lawyer's direction might be viewed differently. In Confiella's copy mode the confirmed, pseudonymised request goes into your own ChatGPT, Claude or Gemini subscription, and the program explains the limits of this beforehand. On a client matter, read in advance what the provider allows itself, and as a lawyer, agree with clients how they should use AI in their case.
Sources
- U.S. District Court for the Southern District of New York: Memorandum, 25 Cr. 503 (JSR)
- Debevoise & Plimpton: SDNY Rules AI-Generated Documents Are Not Protected by Privilege
This article is based on the public sources listed. If you find a mistake, write to us: [email protected]