US – US District Court: AI Public Platforms Chats, Privileges, and Legal Controversies 

The use of AI-supported solutions for professional work, including legal research, tax compliance, preparation for tax disputes, and other closely related advisory practices, is becoming increasingly popular. Generative AI opens the door to an entirely new form of communication between technology and law. 

Taking into consideration the ever-growing usage of Generative AI in the legal practice, in the recent case United States v. Hepner, Judge Jed S. Rakoff of the Southern District of New York issued a bold, clear, and pioneering decision. 

The judge’s opinion expresses that even though the usage of the Generative AI in legal proceedings represents a new frontier, the traditional legal principles still stand before the court. The Court stated that legally relevant information shared with an open(public) AI platform isn’t automatically protected by the client-attorney privilege or work-product doctrine. 

The judge issued an opinion rejecting the criminal defendant’s claim that the dialogue processes in the pre-indictment period with Antropic’s AI platform(Claude) should be as such protected under the client-attorney privilege or work-product doctrine. 

The Judge issued a ruling that could have a much broader impact when it comes to the usage of the public Generative AI platforms, in the context of sharing sensitive legal information, and in the same framework, the attorneys in general should be more “cautious” when it comes to advising their clients “ on with whom and how” they could communicate about “case-worthy” legal information. 

Background

The case includes Bradley Heppner, former CEO and chair of several entities, including one publicly traded company, GWG Holdings, and the federal Southern District court of New York.

 On October 28, 2025, Mr. Heppner was indicted with a variety of criminal counts, including securities and wire fraud. As part of his arrest, FBI agents seized numerous electronic devices, among which were more than 30 documents that recorded dialogue between the defendant and Claude, a generative AI platform operated by Anthropic. 

In these documents, the defendant shared in a dialogue with Claude sensitive legal information, discussing openly his legal approach to “combat” indictment charges. On February 6, 2026, the Government moved for a ruling that the documents generated by Claude aren’t protected by the attorney-client privilege or work-product doctrine. 

Even though material evidence clearly showed that the defendant’s attorney hadn’t in any manner influenced the dialogue between his client and the Antropic’s Claude, the defendant opposed the Government’s motion on the grounds that client-attorney privilege and word-product protection safeguards were breached. 

The defendant’s claim was based on three principal grounds: 

1. The description of the prompts and the connected dialogue with the platform were based on the attorney’s suggestions,
2. The generated documents were extracted from the platform only as part of preparation, getting acquainted with the applicable norms, before meeting with the counsel, and
3. The extracted content was entirely discussed with the attorney 

On February 10, 2026, the Court granted the Government’s motion, ruling that the defendant’s communication with Claude isn’t protected by the client-attorney privilege or work-product doctrines. On February 17, 2026, the judge issued a memorandum in which it has elaborated key principles for deciding in the government’s favour. 

Legal Framework  

Attorney-Client Confidentiality 

The Judge emphasized in his ruling that the attorney-client privilege is of limited scope and protects from pressure to reveal only information that is extracted from the communication between the client and his counselor, which should be kept confidential, as a logical consequence when the individual seeks legal advice. 

As mentioned above, the Judge dismissed the defendant’s argument that the client-attorney privilege wasn’t attended and that the work-product doctrine was not violated. The Judge’s decision to dismiss the complaint was grounded upon the following principles: 

• The Generative AI tool cannot be “treated” as an attorney, and as such, the dialogue between the defendant and the third “person” who isn’t an attorney cannot be treated as a client-attorney relationship, which is the first element to be met, for the applicability of the client-attorney confidentiality rule. 
• There aren’t reasonable grounds for calling upon the existence of privacy in the communication between the defendant and Antropic’s Claude, considering that in the terms&conditions part, the AI Platform indicated that communications and associated personal information may be used for secondary purposes, including training of the AI tool and making disclosures to third parties(such as government agencies).
• The defendant hasn’t communicated with Claude on the counselor’s advice.

Applicability of the work-product doctrine

The work-product doctrine protects documents and other materialized information prepared by attorneys to be shared with adversary parties. The main purpose lies behind the fact of protecting an attorney’s right to prepare for the court, without any sort of “fear” that legally sensitive information stored in this manner could be requested by the adversary. 

In the United States v. Hepner, Judge Jed S. Rakoff of the Southern District of New York, noted that the work-product doctrine couldn’t be called upon a counselor, considering that the defendant communicated with Claude freely without any instructions previously given by this counselor. 

Self-coordinated communication with third parties, as in this case, the AI tool, cannot be under the framework of the work-product doctrine, considering that the main element is missing, and that is the attorney’s work in this “pre-attorney preparation anticipating dialogue”. 

Relevant Insights 

An AI tool isn’t an attorney

Communication with Generative open –  AI platforms cannot be treated from the legal perspective as the dialogue with an accredited counselor. This interpretation should also be taken into account by tax compliance officers when sharing legal/tax-sensitive information with AI tools. 

Confidentiality: Terms&Conditions 

The judge noted that the communication between the defendant and Claude wasn’t confidential. The Anthropic privacy policy gives the platform the right to collect data from users’ input and outputs, including for training of its models, and to disclose the information to third parties. 

Generated content doesn’t fix the controversy 

The defendant argued that he intended to share the documents at a later stage with his attorney. Not-privileged initial communication cannot, on a later stage, become privileged only because it is later shared with the counselor. 

Work-product doctrine 

The documents were generated on the defendant’s own initiative, not at the direction of counsel, and as such, they don’t express the “thought process” of the respective attorney. 

Compliance Recommendations 

Investigate the privacy policy of the open AI-platform before submitting legal or tax-sensitive information, to avoid potential surprises that could arise, if just to name one,  governmental authority requests disclosure of documents. 

• Make a clear distinction between general research and privileged legal analysis 
• Enterprise AI tools are recommended for communication that includes the sharing of sensitive information. 
• Necessity to document attorneys’ participation
• Development of the internal policy for the extent of usage of AI tools

Takeaway

Frequently Asked Questions