Intellectual Property Blog

Claude Isn't Your Lawyer, and the Information You Share with It Isn't Privileged

Generative artificial intelligence (“AI”) tools like Claude, ChatGPT, and Gemini are heavily relied on by people in everyday life. In a lot of ways, they have replaced Google or other general search engines. It is no surprise, therefore, that people now turn to these same tools to answer legal questions. Sometimes those questions are personal and contain facts and information that the person would not share with anyone but an attorney. They may be admissions of guilt or liability or even information that suggests future wrongdoing, and when shared with a person’s attorney, they are privileged. The other side in a legal proceeding does not have any rights to privileged information, but is that the case where the information is submitted to generative AI tools? An answer is beginning to emerge, and it should give all who use these tools pause: prompts and questions submitted to generative AI tools are not privileged. Other parties involved in future litigation and even the police or investigative bodies can likely read everything submitted to those online tools, which are quickly becoming a trap for the unwary. 

The first ruling of this kind came on February 17, 2026, from Judge Jed Rakoff in the case of United States v. Heppner, pending in the Southern District of New York. 25 Cr. 503 (JSR), 2026 WL 436479 (S.D.N.Y. Feb. 17, 2026). The defendant in that case was charged with multiple counts of financial crimes involving his company’s investment clients. He was arrested by the FBI, which issued a search warrant on his home, seizing numerous documents and electronic devices. Many of the seized documents were communications the defendant had with the generative AI platform Claude when it was clear to him he was a target of the investigation after receiving a grand jury subpoena. According to the defendant, he entered prompts into Claude to prepare arguments and defenses to the charges he anticipated. Because his research approximated communications he might have had with an attorney while devising his defense and some of the information he entered he had received from his counsel, he argued that the FBI and the Department of Justice could not see those documents because they are privileged. 

Judge Rakoff ruled the documents are not protected by attorney-client privilege because the communications were not exchanged between the defendant and his attorney, and they were not kept confidential. The attorney-client privilege applies to “a trusting human relationship” with a licensed professional “who owes fiduciary duties and is subject to discipline.” Claude may be able to pass a bar exam, but it is no such “trusting human relationship” with a fiduciary duty to the defendant. Further, Claude’s written policy advises that it (1) collects data on both users’ inputs and its outputs for training purposes and (2) reserves the right to disclose that data with third parties, including governmental authorities. The defendant therefore had no expectation of privacy in his inputs into Claude or its outputs—regardless of whether he printed them out. Further, although it is a closer call, Judge Rakoff ruled that the defendant did not gather the information in furtherance of the case because the defendant’s lawyer did not instruct the Claude inputs and outputs. 

A second form of privilege, known as attorney work product privilege, also did not apply because that privilege only applies materials provided by or at the behest of an attorney in anticipation of litigation. The defendant’s attorney did not himself engage in the Claude prompts or instruct the defendant to do so.  

Why this matters to you: 

It is natural to be curious about what may happen in legal proceedings. It is also natural to wonder if your lawyer’s advice is the best advice. Acting on that curiosity by seeking the counsel of generative AI is dangerous, however. All your inputs and the platforms’ outputs are discoverable in any legal proceeding that is on the horizon; erasing them is not an option. Not only do the rules of discovery explicitly forbid a party from destroying evidence like these prompts, but the generative AI companies could be subpoenaed for the information and produce your inputs and the platforms’ outputs. When in doubt, there is still one old-fashioned method that beats new technology: call your attorney.

Bruce Paul is a Member of McBrayer PLLC, practicing in the firm's Louisville office. His law practice primarily focuses on intellectual property, copyright law, trademarks, commercial and business litigation, employment law, and infringement litigation. Mr. Paul can be reached at bpaul@mcbrayerfirm.com.

Services may be performed by others. This article does not constitute legal advice.