US v. Heppner (S.D.N.Y. 2026): Court says AI chats get no attorney‑client privilege

The ruling

It has been reported that in US v. Heppner, Judge Jed S. Rakoff of the Southern District of New York ruled that communications with generative‑AI tools do not qualify for attorney‑client privilege. The court reasoned that sending confidential legal matters to an AI provider amounts to disclosure to a third party, destroying the confidentiality necessary for the privilege to attach. In short: tell your LLM what you wouldn’t tell a stranger? You’ve likely waived the privilege.

The opinion focused on the practical realities of how AI systems operate — retention, sharing, and potential use of inputs for model training — and held that those features undercut any claim that the communications remained private. It has been reported that the order walked through whether an AI vendor’s terms and technical architecture created a sufficient expectation of confidentiality; the answer, according to the court, was no. Some providers allegedly retain or reuse prompts and outputs, a fact the court treated as fatal to privilege claims.

Why it matters

This is a wrenching moment for law firms and clients who’ve quietly been feeding sensitive work into chatbots to save time. The emotional core here is simple anxiety: are my secrets safe? The ruling signals that, absent contractual or technical guarantees of confidentiality, using public or commercial AI in legal workflows could expose privileged strategies to opposing parties or investigators. Firms must now think twice about where drafts, client facts, and legal theories are composed.

Expect rapid ripples: tighter vendor contracts, new product features promising on‑device or non‑retentive modes, and perhaps a wave of litigation as parties test the boundaries. The decision doesn’t answer every question — e.g., what about private, non‑retaining deployments? — but it sets a clear, conservative baseline. For now, lawyers should treat AI like any third party: assume disclosures risk waiver, and lock down the really sensitive stuff.

Sources: thomsonreuters.com, Hacker News