A federal judge in Manhattan has ruled that documents a former CEO generated using Anthropic's Claude chatbot are not protected by attorney-client privilege and must be turned over to prosecutors. US lawyers across the country are now warning their clients of a new and unintuitive reality. Whatever you type into an AI chatbot can land in front of a grand jury.

The ruling, by US District Judge Jed Rakoff in the case against former Beneficient chairman Bradley Heppner, ordered Heppner to produce 31 documents he created with Claude after receiving a federal grand jury subpoena. Heppner had been the target of an active investigation into alleged fraud at the time he generated them.

Rakoff's reasoning was simple. Privilege requires confidentiality. The moment Heppner shared his thoughts with a third party, in this case Anthropic's Claude, that confidentiality was broken. The chatbot is not his lawyer.

Two Rulings, One Day, Different Answers

The same day Rakoff handed down his ruling, US Magistrate Judge Anthony Patti in Michigan ruled the opposite way in Warner v. Wexford Health. There, a woman representing herself in an employment lawsuit was not required to turn over her ChatGPT conversations about her own case.

The two rulings are not contradictory. They are the opening shots in a body of caselaw that does not yet exist. Rakoff's case involved a represented defendant under criminal investigation. Patti's case involved a pro se litigant in a civil dispute. Different facts produced different outcomes. The harder questions are still to come.

What lawyers are taking from both decisions is a single rule of thumb. Treat any AI chatbot prompt the way you would treat a postcard. If you would not be comfortable seeing it on the back of one, do not type it.

The Industry Has a Confidentiality Problem

OpenAI, Anthropic, and Google all market enterprise versions of their chatbots with explicit data protection guarantees. Consumer versions of the same products do not offer those protections, and most users do not know the difference.

ChatGPT alone now has more than 800 million weekly active users. A meaningful fraction of those users are typing things into the box they would never put in writing anywhere else. They are doing it because the interface feels private. It looks like a text message. It responds like a friend. It is neither.

The companies have not done much to correct the impression. OpenAI's privacy policy makes clear that conversations may be reviewed by humans, retained for safety research, and produced under valid legal process. Most users have not read it.

What Comes Next

The Rakoff ruling will be cited in every criminal case where prosecutors want chatbot conversations from a defendant. It will also be cited in civil discovery, employment disputes, family court, and immigration proceedings. Anthropic itself has been served with multiple subpoenas in the past year. So has OpenAI.

Federal courts have not yet decided the bigger question. Whether chatbot conversations are sometimes covered by a Fifth Amendment self-incrimination protection, the way a defendant's own private notes can be. Several scholars argue they should be. The Department of Justice argues they should not.

Until that fight is resolved, the safest assumption is the one Rakoff put on the record. When you talk to a chatbot, you are talking to a company. Companies can be subpoenaed. Anything you said is now evidence.

Reporting from Reuters via Lee Enterprises, Cliffe Dekker Hofmeyr, and FindSkill.ai.