What happened
On 28 October 2025, a grand jury charged Bradley Heppner with fraud and false statements. During a search of his home, FBI agents seized devices on which they found about 31 documents of chats with Anthropic’s AI service Claude.
According to the defense, Heppner began to use the service after the grand jury subpoena, when it became clear he was the target of the investigation. He did it on his own, without his lawyer’s direction: he entered details he had received from his defense counsel, prepared reports on the defense strategy and possible arguments, and then passed the answers to his lawyer.
On 6 February 2026, the prosecution asked the court to rule that these documents were protected neither by attorney-client privilege nor by the work product doctrine. After a hearing on 10 February, Judge Jed Rakoff granted the motion, and on 17 February he issued a written opinion.
The court called the question one of first impression nationwide. That is why not only criminal defense lawyers read the ruling: it is the first to describe how a court looks at a client’s chats with AI.
The circumstances are typical for any practice. A client under the pressure of an investigation looks for answers in a chatbot on their own, retells their lawyer’s advice to it, and brings the result to the lawyer, believing it is part of the defense. There were few documents, but they held the defense strategy that the defendant had built together with AI.
Why the court refused protection
The court checked the elements of privilege and found none of them.
1. The party to the communication. “Because Claude is not an attorney, that alone disposes of” the privilege claim, Judge Rakoff wrote.
2. Confidentiality. Anthropic’s policy allows it to collect prompts and answers, use them for training and disclose them to third parties, including in litigation. Under such terms, the user has no reasonable expectation of confidentiality.
For lawyers, the second reason matters most: the outcome was decided by the service’s terms, not by the content of the chats. The same terms apply to every user of a public plan, including the lawyer.
3. Purpose. Heppner did not seek legal advice from the service, and the fact that he later passed the answers to his lawyer does not change that.
Note the third reason. Even if a client writes to a chatbot for a future talk with a lawyer, the court looks at whom the client asked for advice. The service is not a lawyer, so a request for advice to it does not create privilege.
The court also rejected the work product argument: the defendant prepared the documents himself, not a lawyer and not at a lawyer’s direction, and they did not reflect defense counsel’s strategy.
The court left one option open. If the client had used AI at the lawyer’s direction, the service might count as the lawyer’s agent under the Kovel doctrine, and the answer might be different.
The court’s position matches what the vendors say themselves. On 25 July 2025, OpenAI’s head Sam Altman said that conversations with ChatGPT have no legal protection like a talk with a lawyer or a doctor, and that the company would have to disclose them if a court required it.
Critique of the ruling
On 23 March 2026, the Harvard Law Review blog published an analysis by Mira Kumar. In the author’s view, the court’s opinion leans toward a categorical rule under which a client’s use of generative AI always stays outside privilege. The author proposes deciding the question on the facts of each case.
The main arguments of the critique:
1. AI is a tool, not a party to a conversation. When a client writes to a lawyer through Gmail, the text also passes through a third party, and nobody asks whether the email service is a lawyer.
2. The “at the lawyer’s direction” requirement will turn into a formality: defense lawyers will start adding a template clause to every engagement.
3. Confidentiality should be judged by the user’s reasonable expectation, not by the vendor’s technical ability to read the text. Service policies change often and are rarely read.
There is also a ruling the other way. In a civil case, a court in the Eastern District of Michigan called AI programs “tools, not persons” and held that materials a party without a lawyer prepared in ChatGPT were work product.
For a reader in Russia, the useful part of the critique is the line of reasoning itself: the court judges not the technology but who receives the text and on what terms. A Russian lawyer answers the same question when choosing a tool for work on matters.
The rule in the US is still taking shape. But it is risky to count on the court adopting the critics’ position: the Heppner opinion stands, and a blog analysis does not overturn it.
Takeaways for lawyers in Russia
Russia has a different regime. Article 8 of the Federal Law on Advocacy protects any information related to legal help to a client, and only the client can release the lawyer from secrecy. Under Article 6, item 5 of the Code of Professional Ethics of Lawyers, secrecy covers information from the client, the documents collected, the content of advice and the whole case file of the lawyer.
The logic of the Heppner case carries over here too. In our reading, secrecy binds the lawyer, while information that the client or the lawyer sent to a public service sits with the vendor under the vendor’s rules, and Article 8 does not protect it there.
What follows for daily work:
1. A lawyer does not enter client information into public AI services without the client’s consent and a check of the processing terms.
2. The client is warned in writing: their own chats about the case with a public chatbot may end up with the other side.
3. If AI is needed for work on a case, it is used at the lawyer’s direction and in the firm’s tool, not in the client’s personal account.
4. The model for work on matters runs on the firm’s server or in a cloud in Russia, and prompts and answers stay in the firm’s log.
5. The agreement with the client describes which AI tools the firm uses and where the data is stored.
A separate question is what the client does with AI on their own. A client who retold the situation to a chatbot before meeting the lawyer has already created text outside the lawyer’s case file. It is worth asking about this at the first meeting.
For in-house legal departments the conclusion is the same, but the basis differs: trade secrets and 152-FZ, Russia’s personal data law, not Article 8. A staff lawyer’s chats with a public service are stored by the vendor under the vendor’s rules.
The rules of secrecy for neural networks and the position of the Federal Chamber of Lawyers are collected in the article on attorney secrecy and AI, and consent wording in the article on client consent to AI.