If you have recently asked Claude or ChatGPT about sensitive matters—bankruptcy strategies, tax arrangements, or how to handle a dispute with your boss—more than a dozen top U.S. law firms issued client alerts in mid-April: those conversations could now be used as court evidence.
This is not a scare tactic. It has already happened.
Two Rulings, Two Directions
Heppner case, U.S. District Court for the Southern District of New York, February 17, 2026.
Defendant Bradley Heppner, former chairman of the bankrupt GWG Holdings, is charged with securities fraud. Heppner used Claude to prepare a set of documents related to the case, claiming they were for his legal team. Defense attorneys argued that the AI-generated content was protected by the attorney work-product doctrine.
Judge Jed Rakoff ruled directly—it is not protected.
Three reasons:
- The conversations were not created by an attorney; no licensed lawyer was involved.
- Anthropic's privacy policy allows disclosure of user data to third parties.
- Heppner consulted the AI on his own, not at the direction of his attorney.
In other words: you can treat AI as a paralegal, but the law does not recognize it as one.
Warner v. Gilbarco case, Michigan, same day.
A pro se plaintiff in an employment discrimination case submitted her conversations with ChatGPT about the case to the court. The court ruled the opposite: the chats are protected under the work-product doctrine.
The reasoning: AI is a "tool, not a person"—just as using Excel to calculate evidence does not mean Excel's terms constitute third-party disclosure.
Two rulings, same day, opposite directions. That is the problem—the same AI conversation could have completely different outcomes depending on which court hears the case.
What Really Set This Off
In January 2026, another Southern District of New York judge, Sidney Stein, ordered OpenAI to hand over 20 million de-identified ChatGPT conversation logs to The New York Times and other plaintiffs as evidence discovery in a copyright lawsuit.
OpenAI argued this violated user privacy. The objection was overruled.
Those 20 million conversations were not from a suspect—they were randomly sampled from ordinary users.
Then in April, more than a dozen top U.S. law firms—not small firms, but firms like Skadden and Latham—simultaneously sent alerts to clients:
- AI conversations may be subject to discovery requests under Rule 26.
- Companion apps (Character.AI, Replika, Nomi) already have privacy policies that allow disclosure to law enforcement.
- Investigators do not need to subpoena the platform directly; they can ask the user to export their own chat logs and submit them.
The Stored Communications Act restricts subpoenas to platforms, but it does not restrict a user from exporting their own data.
What Users Should Do
In practice, there is not much you can do.
- Assume that everything you discuss with an AI could theoretically be obtained by a court.
- Attorney-client privilege protects conversations with a licensed attorney, not with an AI—even if the AI gives better answers than the attorney.
- Work-product protection is applied inconsistently across courts. Do not gamble on it.
- The line in privacy policies that says "we may disclose to third parties" is not decorative. It is enforceable.
OpenAI's paid ChatGPT tier has a "do not retain conversations" option, but that is a product feature, not a legal exemption. When a court order arrives, the meaning of "do not retain" will be reinterpreted.
Industry Reaction
Character.AI was already ruled a "product, not protected speech" last May—the First Amendment defense is gone. Anthropic's policies are similar. OpenAI's policies are more aggressive—its enterprise customers can negotiate zero retention, but ordinary users cannot.
The next development to watch is the appeal in the Heppner case. If the Second Circuit upholds the ruling, the rest of the country will likely follow. If it is overturned, the fight starts over.
But even if the appeal succeeds, those 20 million logs are already in the hands of the NYT legal team.
AI conversations are no longer a private matter between a person and a machine. They are now potential evidence. A dozen law firms sent warnings in the same week—not to scare you, but because they are already handling these issues for their clients.
Sources: Federal Judge Rules AI Chatbot Conversations Can Be Used as Evidence in Court (Technology.org); OpenAI Loses Privacy Gambit: 20 Million ChatGPT Logs Likely Headed to Copyright Plaintiffs (National Law Review); CocoLoop, AI Chats Not Privileged: Federal Court Ruling (The Outpost); Federal Courts Just Made Your AI Chats Court Evidence in April 2026 (RoboRhythms)