Thinking about using ChatGPT, Claude, or Perplexity to collect your thoughts for an email to your lawyers? Don’t assume your chat will stay confidential.
A federal judge ruled on Tuesday that prosecutors could access Claude chat transcripts generated by Brad Heppner, a finance startup founder accused of defrauding a company out of $150 million.
The chats occurred after Heppner received a subpoena, hired lawyers, and learned that he was a target of prosecutors, his lawyer said in court.
Heppner, who helped start the finance firm Beneficient, was arrested last year and charged with wire and securities fraud for conduct that allegedly led to the downfall of GWG Holdings.
Investigators seized “dozens of electronic devices” when they arrested Heppner at his Dallas mansion, prosecutors said, and Heppner’s lawyers have insisted that 31 chats with Anthropic’s Claude bot on those devices are privileged.
“Mr. Heppner — using an AI tool — prepared reports that outlined defense strategy, that outlined what he might argue with respect to the facts and the law that we anticipated that the government might be charging,” his lawyer said.
“The purpose of his preparing these reports was to share them with us so that he could discuss defense strategy with us.”
Even though Heppner had privileged conversations with his lawyers, Judge Jed Rakoff said he “disclosed it to a third-party, in effect, AI, which had an express provision that what was submitted was not confidential,” according to a transcript of the hearing.
The government noted that Claude’s privacy policy specified that chats could be disclosed. Prosecutors also said that the chats couldn’t be protected by the “work product privilege,” which can guard materials prepared at a lawyer’s direction, because Heppner’s lawyers didn’t ask him to use Claude.
The decision has lawyers buzzing.
“My gut reaction is that the decision is directionally correct,” Moish Peltz, an attorney whose post about the decision ricocheted around X, told Business Insider. “There’s a lot of materials that should be kept as privileged that people are putting into AI.”
The proliferation of chatbots where people are inputting sensitive legal information, another wrote, has created “a discovery nightmare.”
Noah Bunzl, an employment lawyer, told Business Insider that people might find it “somewhat shocking” that their legal confidences could be lost by sharing them with a chatbot.
The case isn’t the first where an executive’s use of a chatbot was the subject of legal debate.
In November, PC Gamer reported on a dispute involving the acquisition of a video-game company, where a company official’s use of ChatGPT to try to avoid paying an earn-out was mentioned in court records.
And after The New York Times sued OpenAI for allegedly violating its news article copyrights, a judge required OpenAI to retain millions of chat logs to potentially review them for copyright infringement.
Bunzl said he has noticed that in civil discovery, lawyers are increasingly asking for their adversary’s AI chats. It’s “a whole other world of discoverable information,” he said.
Still, attorneys at the law firm Debevoise & Plimpton, who analyzed the Heppner decision, said it was the first they were aware of where someone’s use of an AI tool may have resulted in “a loss of privilege” over privileged material. They said courts may view businesses’ use of purpose-built AI tools differently.
Arlo Devlin Brown, a white-collar defense lawyer, told Business Insider he thought AI models could potentially improve attorney-client information. But given the ambiguity in the law, people have to be vigilant.
“Until the law has been clarified, lawyers should caution their clients that inputting otherwise privileged information into an AI tool could risk exposure in litigation,” he said in an email.
Representatives for the US Attorney’s Office for the Southern District of New York and Anthropic, and lawyers for Heppner, didn’t immediately reply to requests for comment.
