U.S. Court Rules AI Conversations Not Protected by Attorney-Client Privilege

In November 2025, the U.S. Securities and Exchange Commission (SEC) and the Department of Justice (DOJ) launched investigations into executives of multiple publicly traded companies. Bradley Heppner, an executive of GWG Holdings, Inc. and several other companies, was formally indicted on charges of securities fraud, wire fraud, and conspiracy. The indictment alleged that he defrauded investors of over $150 million through false statements and undisclosed related-party transactions.

During the arrest operation in November 2025, the Federal Bureau of Investigation (FBI) executed a search warrant at Heppner’s home and seized numerous electronic devices. While reviewing the evidence, prosecutors discovered approximately 31 documents documenting text exchanges between Heppner and a generative AI platform named “Claude” (developed by Anthropic).

Heppner’s attorney argued that these AI conversation records (hereinafter referred to as the “AI Files”) should be protected by the attorney-client privilege and the work product doctrine. Because when conversing with the AI, Heppner input information obtained from his attorney, requested the AI to assist in analyzing defense strategies and organizing legal arguments, and these records were created for subsequent communication with his attorney and trial preparation; moreover, Heppner had actually shared the AI-generated content with his defense team.

On February 6, 2026, the prosecution filed a motion with the court, seeking to compel a finding that the AI Files are not privileged and should be disclosed to the government as evidence. The parties engaged in heated debate on this issue during a pretrial conference on February 10. Presiding Judge Jed S. Rakoff rejected the defendant’s argument in court and, on February 26, issued a detailed legal memorandum explaining the reasoning for his ruling.
The judge noted that the application of attorney-client privilege requires the simultaneous satisfaction of three elements:

First, the privileged communication must occur between a “client” and an “attorney.” Claude is a machine, not a licensed legal professional bound by professional ethics. There is no “confidential relationship” or “fiduciary duty” between a user and an AI.

Second, according to Anthropic’s public privacy policy, user inputs are collected for training the AI model and may be disclosed to third parties under certain circumstances (e.g., at the request of government regulatory authorities or in legal disputes). The court held that a user who voluntarily provides information to such a public platform waives a reasonable expectation of privacy, and thus the conversations themselves lack confidentiality.

Third, Heppner’s purpose in communicating with Claude was to seek assistance from the AI, rather than to consult an attorney. The AI platform itself explicitly states, “I am not an attorney and do not provide formal legal advice.” Even if the user later forwards the content to an attorney, it cannot cause the initial non-privileged communication to acquire privilege “retroactively.”

The judge further noted that the AI Files are not protected by the work product doctrine. This is because the work product doctrine is designed to protect an attorney’s thought processes, providing a private space for attorneys to prepare for litigation. The doctrine generally applies to materials prepared by an attorney or their agent. The court found that Heppner admitted his attorney had not instructed him to use Claude for searches. He interacted with the AI entirely voluntarily and independently, and did not act as the attorney’s “agent.” Although the AI-generated content may have influenced the attorney’s subsequent strategy, these files themselves reflect the AI’s output and do not directly reflect the defense attorney’s mental processes or strategy.

This case is considered the first federal criminal case in the United States and globally involving whether conversations with a generative AI are protected by attorney-client privilege. Currently, Heppner’s case is scheduled for trial on April 6, 2026. This court ruling means the prosecution will have the right to review these 31 AI conversation records, which may serve as prosecution evidence, posing a significant challenge to the defendant’s defense strategy.