AI in Legal Practice: What Heppner v. US Means for Lawyers, Mediators, and Arbitrators

For the sake of transparency and self-protection, let me start by saying I am not an attorney, and I am certainly not offering legal advice.

However, this should unsettle every legal professional (attorney, mediator, arbitrator, etc.)  reading this. In February 2026, a federal judge in the Southern District of New York handed down a ruling that many of us in the legal and ADR community had been quietly dreading. Not because it was surprising, exactly. More because it confirmed what some of us already suspected: that the casual use of AI tools in legal practice has been creating exposure we cannot see and cannot easily undo. 

The case is United States v. Heppner (25-cr-00503-JSR). Judge Jed Rakoff, one of the sharper legal minds on the federal bench, described it as “a question of first impression nationwide.” He wasn’t exaggerating.

What Actually Happened

Bradley Heppner, a financial executive facing federal fraud charges, used Anthropic’s publicly available version of Claude to prepare defense strategy materials. He later shared those materials with his attorneys. Sounds reasonable, right? Orderliness was kind of his whole personality. He was trying to be a good client.

The FBI seized his devices during a search warrant. Thirty-one documents, AI generated, all containing legal strategy. His lawyers argued privilege. The government disagreed. Judge Rakoff sided with the government for three distinct reasons that every practitioner needs to internalize.

1. Claude is not an attorney. Communications with an AI tool are not communications with counsel!
   
2. Heppner had used the free, public version of the platform. Anthropic’s own privacy policy explicitly states that user inputs may be used for model training and disclosed to third parties, including government regulatory authorities. There was no reasonable expectation of confidentiality.
   
3. The materials were not created at the direction of counsel. Heppner acted on his own initiative, which meant the Kovel doctrine (which can extend privilege to certain non-attorney agents acting under lawyer direction) simply did not apply.

The court ordered disclosure to the prosecutor. The materials were not protected.

Why is this a major problem?

I’ve thought about the infrastructure reality beneath this ruling, and honestly, the privilege question is almost the smaller concern. Think about what it means that an AI company’s servers contain a log of everything you typed and what this means in terms of confidentiality & privacy. Every contract clause you asked about, litigation strategy you explored, or settlement range (ZOPA) you tested.

Well, guess what? Those logs exist. They are stored on servers in locations you do not know, across jurisdictions whose laws you have not read, potentially reviewed by human employees whose vetting you cannot audit. Under U.S. law, the Stored Communications Act (18 U.S.C. § 2703) establishes the baseline legal process by which the government may compel disclosure of stored electronic communications from service providers, including communications stored in contemporary cloud environments. Law enforcement has used it. In a separate and equally instructive case, the CEO of Krafton used ChatGPT to, as the Delaware court record describes it, brainstorm ways to avoid paying $250 million contractual earn-out to game developers. He deleted the conversations. They surfaced anyway as central evidence in litigation. Deletion from your interface does not mean deletion from their servers! For lawyers, this is a competence and confidentiality problem. For mediators and arbitrators, it is arguably more serious. Your confidentiality obligations run to all parties simultaneously, not just your own client. Feeding case details, financial terms, or party disclosures into a public AI tool may expose one party’s private caucus information in ways that no attorney in the room would have any right to do.

The Silicon Valley Arbitration and Mediation Center’s Guidelines on the Use of AI in Arbitration (April 2024) are unambiguous: an arbitrator should not input any information about an arbitration into an open AI system.  While the guidelines are strong, they are designed as a “principle-based framework” rather than binding rules of law, although they guide ethical conduct. The guideline on confidentiality (Guideline 2) states that participants “should not submit confidential information to any AI tool without appropriate vetting and authorization”. It advises against using public AI tools (like free ChatGPT) that store data, effectively making it a prohibition on unsecured use.

What the Court Left Open (and Why It Matters)

Here is something the commentary has mostly missed. Heppner is not a ban on AI in legal work. Judge Rakoff himself noted that had counsel directed Heppner to use Claude, the outcome might have been different. Claude might have functioned, in his words, “in a manner akin to a highly trained professional who may act as a lawyer’s agent within protecting the attorney-client privilege.”

A separate federal decision was issued the same week, Warner v. Gilbarco (E.D. Mich., Feb. 10, 2026), upheld work product protection over a pro se plaintiff’s AI-assisted materials, reasoning that AI platforms are “tools, not persons” and that disclosure to software is not disclosure to an adversary.

So, confusion reigns, and the law is genuinely unsettled. What is settled is the risk framework. Consumer AI tools, used without attorney direction, without confidentiality protection, and without documented workflow supervision, will not be protected. Enterprise tools with zero-retention contractual guarantees, used under attorney direction with documented protocols, have a much stronger case. The difference between those two scenarios is not AI. It is everything around the AI.

Practical Actions: What to Do, What to Avoid

Stop doing these things:

• Do not paste client names, matter details, financial terms, or any identifying information (Pii) into any consumer AI platform. This includes the free tiers of Claude, ChatGPT, Gemini, and similar tools. It includes the version you access through your browser without a corporate agreement. ABA Formal Opinion 512 (July 2024) is explicit: without explicit client consent and robust security guarantees from the AI provider, using consumer AI tools for client matters is extremely risky.
  
• Do not assume that deleting a conversation protects you. The Krafton case should have ended that assumption. Logs persist. Backups exist. Training datasets may already contain your input.
  
• Do not let junior staff or clients use AI tools on matters without a written policy governing that use. A 2026 survey found that 75% of lawyers were using AI but only 25% had received formal training on the ethical implications. That gap is a liability gap.

Start doing these things now:

• Update your engagement letters. They need to address AI use explicitly, yours and your client’s. Warn clients that communications with public AI platforms are not confidential and are potentially discoverable.
  
• If you use AI for document analysis, implement a de-identification step before any document leaves your local environment. Tools like PII Anomalyzer (https://azdecisionscience.com/pii-anomalyzer) are built specifically for this workflow. The software runs entirely on your local machine, MacOS or Windows, with no cloud access. It identifies and replaces personal identifying information with semantically consistent tokens that preserve the analytical utility of the document. The real data never leaves your device. The mapping that allows you to re-identify outputs is stored locally, under your control. When you submit the de-identified version to an AI tool, the AI has no identifiable client data to store, subpoena, or inadvertently train on. This is not a workaround. It is the architecturally correct solution to the problem Heppner created.
  
• For attorneys and their clients, document your AI use. Keep a record of what tool was used, for what purpose, that no confidential matter details were submitted in identifiable form, and that the workflow was conducted under attorney supervision. That documentation becomes your defense if a privilege challenge arises.
  
• For arbitrators specifically, disclose AI use to all parties at the outset of any proceeding. Get written consent. Be specific, as Daniel Gonzalez of Hogan Lovells noted at the 2025 AAA Future Dispute Resolution Conference, generic references to “transparency” are not enough. Parties need to know what tools are in play and how confidentiality is being handled.

Conclusion

The legal profession has navigated technological disruption before. Email was once a confidentiality concern. Electronic discovery reshaped litigation practice entirely. Each time, the profession adapted, not by avoiding the technology, but by building protocols that made responsible use possible.

AI is no different. The practitioners who will be most exposed are not those who use it, but those who use it carelessly, without understanding what happens to the information they type, and without the procedural guardrails that distinguish professional use from consumer use.

The Heppner case is not the end of AI in legal practice. It is the beginning of a period in which the standards for responsible use are being defined in real time, in courtrooms, ethics opinions, and institutional guidelines. The question is whether you are ahead of that standard or behind it.

The author is not providing legal advice. Practitioners should consult applicable state bar ethics opinions and institutional guidelines governing their specific practice area and jurisdiction.

Robert Bergman

Robert Bergman with Next Level Mediation provides full mediation services - including proprietary and confidential Decision Science (DS) analysis that assists each party in understanding their true litigation priorities as aligned with their business objectives. Each party receives a one-time user license to access our exclusive DS Application Cloud. We… MORE