Court Rules Client’s AI Communications Were Not Protected by Attorney-Client Privilege or Attorney Work Product Doctrine

08.10.2026

Generative artificial intelligence can be useful in legal matters, but it also raises important questions about confidentiality. On February 17, 2026, the United States District Court for the Southern District of New York addressed those questions in United States v. Heppner. The decision does not control California courts, but it offers practical guidance for clients and attorneys who want to protect confidential communications when using generative AI.

Heppner involved Bradley Heppner, who was charged with fraud and making false statements. On November 4, 2025, the FBI seized electronic devices and documents containing Heppner’s communications with “Claude,” a generative AI platform operated by Anthropic, PBC. Heppner had used Claude to help prepare his defense strategy.

Heppner’s attorneys had not told him to use Claude. Even so, Heppner argued that his communications with Claude should be protected by the attorney-client privilege because they included information Heppner had learned from his attorneys, were created so he could consult with his attorneys, and were later shared with them.

The court disagreed. It explained that the attorney-client privilege applies only when a communication: (1) is between a client and an attorney, (2) is intended to be confidential and is actually kept confidential, and (3) is made to obtain or provide legal advice. The court found that Heppner’s communications with Claude did not meet those requirements.

First, Claude was not Heppner’s attorney. The court stated that “Claude is not an attorney,” and that the privilege depends on a human relationship in which one person owes a legal duty to another.

Second, Heppner had been told through Claude’s user agreement that his communications were not confidential. The court noted that Claude users agree to Anthropic’s privacy policy, which allows Anthropic to collect prompts and outputs and disclose that data to third parties, including government agencies.

Third, the court found that Heppner was not using Claude to obtain legal advice from Claude itself. Instead, he was using Claude to help him prepare for discussions with his attorneys. The court concluded that this was not enough for the privilege to apply.

The court suggested the result might have been different if Heppner had used Claude at his attorney’s direction. In that situation, Claude might have functioned more like an outside professional assisting the attorney, which can sometimes fall within the attorney-client privilege.

The court also held that the attorney work product doctrine did not apply. The communications were not prepared by, or at the request of, Heppner’s attorneys and did not reflect their legal strategy. The court did not decide whether the doctrine would apply if an attorney had directed the client’s use of Claude.

California law protects confidential attorney-client communications and writings that reflect an attorney’s “impressions, conclusions, opinions, legal research, or theories.” Although Heppner is not binding in California, it identifies issues California clients and attorneys should consider when using generative AI in legal matters.

A key concern is how generative AI platforms handle user data. Anyone using AI for legal matters should avoid platforms that allow broad third-party access to prompts or outputs. More secure platforms may encrypt user data, limit outside access, and provide protected storage. These features may help show that AI prompts and outputs were intended to remain confidential.

This concern also applies outside the litigation context. For example, people handling confidential information in personnel investigations or other sensitive matters should not enter that information into an AI platform unless they understand who may access it.

Clients and attorneys should also consider who is directing the AI use. Heppner used Claude on his own, without direction from his attorneys. For that reason, the court found that Claude was not acting as his attorneys’ agent, and the materials Claude produced were not attorney work product.

If clients and attorneys decide to use generative AI in a legal matter, they should avoid the circumstances present in Heppner. Clients should not use generative AI for a legal matter on their own before consulting counsel. When AI is appropriate, it is safer for the attorney to direct or perform the use. Under Heppner and California privilege principles, AI communications are more likely to be protected when an attorney uses AI consistent with applicable ethical rules to conduct legal research or form legal advice for a client.

The law on privilege and generative AI is still developing, and California courts may reach different conclusions than the Heppner court. Still, Heppner highlights practical steps that may help clients protect confidential information when using generative AI.

AALRR will continue to monitor legal and policy developments in this area. If you have questions about this Alert, please contact the authors or your usual AALRR counsel.