Your AI Chat History Is Now a Legal Liability | Blog

A federal judge in New York just handed down a ruling that should make every professional who has ever typed sensitive details into an AI chatbot sit up straight. AI chat logs are not protected by attorney-client privilege. They can be subpoenaed. They can be used against you. And the courts are only getting started.

The Case That Changed Everything

In United States v. Heppner (S.D.N.Y., February 2026), Judge Jed Rakoff ruled that 31 documents generated by Anthropic’s Claude — created by the defendant, a former financial CEO accused of defrauding investors of $300 million — were fully discoverable by prosecutors. The defendant had fed details from his lawyers directly into Claude and used the AI’s outputs to build a defense strategy. He argued those documents were protected as attorney work product.

Rakoff dismantled the argument in three moves. First, Claude is not an attorney. Second, Anthropic’s own privacy policy explicitly warns that user data can be shared in legal proceedings. Third, the defendant created the documents on his own initiative, not at counsel’s direction — so they couldn’t reflect the legal strategy privilege is designed to protect.

The ruling is reportedly the first of its kind in the country. It won’t be the last.

Courts Are Not Aligning

If you’re hoping for a clean legal principle to emerge, don’t. Courts are already splitting. In Warner v. Gilbarco (2026), a federal court reached the opposite conclusion for a pro se plaintiff: her ChatGPT research logs were considered work product because they were prepared in anticipation of litigation. The distinction matters — represented parties who consult AI appear to be on much shakier ground than self-represented individuals doing their own research.

Meanwhile, in a separate case, Judge Sidney Stein upheld an order requiring OpenAI to produce 20 million ChatGPT conversation logs as evidence in the NYT copyright litigation. The scale is staggering. AI conversations are no longer ephemeral. They are durable, discoverable, and increasingly attractive to opposing counsel.

The Advice Confusion Problem

There’s a parallel and deeply uncomfortable issue underneath the privilege question: people are treating AI responses as legal advice, and judges are not playing along. Lawyers are now warning clients — in formal client alerts from firms like Fisher Phillips, Dorsey, and Freshfields — that AI tools do not form confidential relationships, do not hold attorney-client duties, and cannot legally advise anyone.

The concern isn’t hypothetical. Individuals facing employment disputes, custody battles, and financial fraud investigations have been documenting their reasoning inside AI chat sessions, assuming those conversations carry the same weight as notes with their attorney. They do not. When that chat log surfaces in discovery, it often undermines the very defense it was meant to support.

Fully Automated Products Are the Next Frontier

While the chat privilege battle plays out, a second legal wave is building around companies that have deployed fully AI-powered public-facing applications. The cases arriving in 2026 share a common thread: opaque automated systems making consequential decisions — insurance denials, hiring screens, content moderation — without meaningful human review.

Cigna, Humana, and UnitedHealth have each faced lawsuits alleging their AI systems wrongfully denied medical claims at scale; in one filing, an algorithm rejected over 300,000 claims in two months. In January 2026, a proposed class action against Eightfold AI alleged the hiring platform compiled sensitive personal data on job applicants without consent, violating federal and state law. The EU’s Product Liability Directive has now codified what many expected: AI software is a “product,” and developers, deployers, and anyone who substantially modifies an AI system can be held strictly liable for damages it causes.

The key problem is what K&L Gates described bluntly: AI systems “fail silently.” They issue decisions without traceable reasoning, without proof, and without a verifiable path back to why an outcome occurred. Courts are increasingly unwilling to accept probabilistic confidence as a legal standard of reasonableness.

What Needs to Change

The legal system is moving faster on this than most technologists expected, and slower than most lawyers hoped. For individuals, the practical reality is stark: anything you type into a consumer AI tool is potentially discoverable. Treat it like email, not like a private journal. For companies deploying AI products, the days of hiding behind terms of service and black-box outputs are numbered. Courts want a traceable chain of accountability. If your product makes a decision that harms someone, you need to be able to explain exactly how and why — and “the model said so” will not hold up.

The rulings are early and inconsistent, but the direction is clear. The courtroom is catching up to the deployment curve, and it is arriving with subpoenas.

AI Disclosure

This document is drafted by an AI skill and is provided for informational and governance support purposes only. It does not constitute legal advice or a formal compliance determination. Do not publish or rely on this notice as a substitute for review by qualified legal counsel or a licensed compliance professional with jurisdiction-specific expertise.