Most lawyers just watched this decision scroll past on X and reacted emotionally. Some said it was nonsense. Some said it would be overturned. Some said judges are protecting their own industry. Some said "just use local models." None of that changes the core issue.
A federal judge ruled that 31 documents a defendant generated using an AI tool and later shared with his lawyers are not protected by attorney-client privilege or work product doctrine.
The case is United States v. Heppner. Judge Rakoff of the Southern District of New York ruled from the bench on 10 February 2026. The defendant had prepared the documents with Anthropic's Claude on his own initiative after engaging counsel, then sent them to his lawyers. The court found he had disclosed them to a third-party tool whose terms said submissions were not confidential, and that the materials were not prepared by or at the direction of counsel (Debevoise & Plimpton, February 2026). Commentators suggest enterprise tools that don't train on inputs and keep them confidential may be treated differently. The ruling doesn't settle that.
This is not a philosophical debate. It is a structural warning to the legal profession.
Your AI Conversations Are Not Privileged
And the Court Just Confirmed It.
The court's reasoning was not dramatic. It was doctrinal.
Attorney-client privilege requires: a communication, between attorney and client, for the purpose of obtaining legal advice, and made in confidence.
An AI tool is not an attorney. It has no law license. It owes no duty of loyalty. It owes no duty of confidentiality. Its terms explicitly disclaim any attorney-client relationship.
If you input your legal strategy into a commercial AI platform, you are communicating with a third party. That destroys privilege.
It does not matter that the interface feels conversational. It does not matter that it feels like advice. It does not matter that you later forwarded the output to your lawyer. You cannot retroactively create privilege by sending non-privileged material to counsel. Courts have been clear on that for decades. The only difference now is that the "third party" happens to be AI.
The Privacy Policy Problem No One Reads
What made the situation worse for the defendant was the AI provider's own privacy policy.
At the time of use, the provider expressly reserved the right to collect prompts, retain outputs, use data for training, and disclose information to governmental authorities and third parties.
That clause alone undermines any claim of a reasonable expectation of confidentiality. Privilege requires confidentiality. If the platform reserves the right to disclose your data, your expectation of confidentiality collapses.
The user experience may feel private. The legal reality is not. Unless you have negotiated an enterprise agreement that changes those terms, you are typing sensitive legal information into a third-party commercial system that retains data and reserves broad rights. That is not privilege. That is disclosure.
The Dangerous Wrinkle: When Counsel Becomes a Witness
The judge also flagged something more serious. The defendant reportedly fed information received from his own attorneys into the AI tool. If prosecutors attempt to use those AI-generated documents at trial, defense counsel could become a fact witness.
That creates disqualification risk, ethical complications, evidentiary instability, and potential mistrial exposure. Winning or losing the privilege motion does not simplify what comes next. AI is not just a confidentiality issue. It is a litigation risk multiplier if used improperly.
Why the Public Reaction Misses the Point
The reactions online were predictable: "The system is protecting itself." "This is why confidential AI is necessary." "This will be overturned." "Lawyers are afraid of losing control."
The reality is far less dramatic. The ruling is doctrinally consistent with long-standing privilege law. What changed is not the doctrine. What changed is user behavior.
People experience AI as a sounding board, a silent advisor, a personal research assistant. But legally, it is a third-party platform. That psychological gap is the real problem.
The Core Risk for Law Firms and General Counsel
If your clients are using public AI tools to summarize your advice, stress-test legal strategy, draft internal risk memos, prepare for litigation, or brainstorm negotiation tactics, those prompts may be discoverable.
Every prompt is a potential disclosure. Every output is a potentially discoverable document.
If you are not proactively advising clients on this, you are already behind.
What Lawyers Must Do Immediately
1. Update Engagement Letters
Explicitly state: any information input into public AI platforms may not be privileged and may be discoverable. Do not assume clients understand this distinction. They do not.
2. Address It During Onboarding
Make it part of your intake conversation. Clients need to understand that AI chat logs are not the same as confidential communications with counsel.
3. Do Not Rely on Hope
Saying "just don't use AI" is unrealistic. Clients will use it. The only serious response is to design safer infrastructure.
The Architectural Solution: AI That Keeps Privileged Material Under Your Control
The long-term answer is not prohibition. It is controlled integration.
If AI is going to be used in legal matters, it must operate within the attorney-client relationship, under lawyer supervision, inside secure firm-controlled environments, with defined governance, with auditability, with no training on client data, and with jurisdiction-aware data handling.
This is not about marketing language. This is about professional responsibility.
In our Legal AI Workshop materials, we emphasize that any legal professional using AI must ensure four obligations are met: Disclosure, Competence, Confidentiality, and Oversight. These are not optional. They are structural.
Vyzkoušejte HAQQ AI zdarma
Zažijte sepisování a rešerše s podporou AI
Why Generic AI Is Not Built for Legal Privilege
Public AI platforms are designed for scale. They are built to optimize performance, improve models, aggregate data, and serve millions of users. They are not designed to replicate the attorney-client privilege structure.
That does not make them malicious. It makes them commercially rational. But legal privilege is not a commercial concept. It is a professional one.
There is a difference between "AI that generates text" and "AI embedded inside a law firm's operating system." That difference determines risk.
If the Provider Keeps It, a Court Can Reach It
On 2 December 2025, in The New York Times v. OpenAI, the court ordered OpenAI to produce 20 million ChatGPT logs after denying OpenAI's motion for reconsideration (Case 1:23-cv-11195, Doc. 1049; Digital Policy Alert). Retention is the point. Whatever a provider keeps, a court can order it to hand over.
What Changes in the Gulf
No GCC court has ruled on AI and privilege yet. Two duties already apply.
- DIFC: Practical Guidance Note No. 2 of 2023 tells practitioners not to disclose a client's confidential information to an AI tool without the client's consent, citing paragraph 19 of the DIFC Courts' Code of Conduct.
- Onshore UAE: Federal Law No. 34 of 2022 broadened lawyers' confidentiality obligations (Norton Rose Fulbright).
The practical rule is the same in both: get written client consent before confidential material goes into any AI tool, and prefer tools that don't train on your inputs. The full map of UAE AI rules is in Onshore, DIFC, ADGM.
The HAQQ Legal AI Perspective
At HAQQ Legal AI, we have been clear: AI is not a chatbot layered on top of legal work. It must be integrated into the legal operating system. It must function inside structured workspaces tied to matters, clients, permissions, roles, audit trails, and jurisdictional controls.
It must be governed. It must be supervised. And it must operate under the lawyer's responsibility, not outside it.
Legal AI without structural safeguards is not innovation. It is exposure.
The Bigger Picture: This Will Not Be the Last Case
This decision is only the beginning. We will likely see future litigation addressing local models, on-premise deployments, enterprise agreements, government subpoenas, cross-border data transfers, health and mental health AI use, and corporate compliance contexts.
The legal profession is entering a new phase. The question is no longer "Can AI help draft?" The question is "Where does AI sit in the privilege structure?" Firms that answer that correctly will protect their clients and themselves. Firms that ignore it will eventually confront it in discovery.
Final Take
If you are a managing partner, a general counsel, a litigation strategist, or a compliance officer, you cannot treat AI as a casual research tool anymore.
This ruling makes one thing clear: AI conversations outside the attorney-client relationship are not privileged.
The illusion of privacy is not privacy. The feeling of confidentiality is not confidentiality. If AI is going to sit at the table in legal strategy, privileged material has to stay under your control. Anything else is risk disguised as convenience.
