Consumer AI can destroy privilege: what United States v. Heppner means for solo litigators

Seventy-one percent of solo attorneys now use AI in their practice (Clio's 2026 Legal Trends Report), and a large share of that is consumer-grade chat — ChatGPT or Claude opened in a browser tab, on a personal account, with no firm controls around it. A recent ruling out of the Southern District of New York is the clearest warning yet about why that habit can be expensive. If you've wondered whether AI attorney-client privilege protects what you type into a chatbot — is ChatGPT privileged? — this opinion gives a concrete answer on its facts.

What happened in Heppner

In United States v. Heppner, No. 25-cr-00503 (JSR) (S.D.N.Y. Feb. 17, 2026) (Rakoff, J.), the court held that roughly 31 documents the defendant, Bradley Heppner, created using a consumer-tier AI assistant were notprotected by the attorney-client privilege or the work-product doctrine. The documents were produced in a written Memorandum on February 17, 2026 (the court had announced the result from the bench on February 10). What made them vulnerable was not the model — Heppner used Anthropic's Claude on the consumer tier — but the posture: he generated them on his own initiative, for his own purposes, without any direction from counsel. You can read the docket on CourtListener.

Why the court ruled that way — three reasons

The reasoning was narrow and, for anyone using AI in a matter, genuinely instructive. Three threads ran through it:

• The AI isn't a lawyer. The privilege protects confidential communications between a client and an attorney for the purpose of obtaining legal advice. Typing a question into a chatbot creates no such relationship — there is no lawyer on the other end, so there is no privileged communication to protect.
• No confidentiality, because of the consumer terms.The consumer product's privacy policy permitted the provider to use inputs to improve its models and to disclose them to third parties, including in response to governmental process. Privilege depends on an expectation of confidentiality; the court found that the consumer terms defeated it. This is the heart of the consumer AI privilege waiverproblem — the platform's own terms undercut the secrecy the privilege requires.
• No attorney directed the use.Heppner ran the tool himself, for himself. There was no lawyer steering the work, so there was no attorney-client communication and nothing prepared at counsel's behest.

The court was careful to leave a door open — a Kovel-style caveat. Had an attorney directed the use of the tool — the way a lawyer directs an accountant or other agent retained to assist in rendering legal advice — and had the work been done on a confidential platform, the result might have come out differently. (Commentators have analogized the situation to Da Silva Moore's early treatment of technology-assisted review; that analogy is the bar's, not something the Heppner court itself relied on.) The opinion did not bless or condemn any particular product. It drew the line at consumer-tier use with no counsel direction.

The takeaway

The lesson is not “don't use AI.” It is that privilege follows the attorney's role and a confidential, no-training platform — not the sophistication of the tool. A frontier model on a consumer account does not earn privilege; a humble workflow that a lawyer directs, on a platform contractually barred from training on the data, has a far better claim to it. And a point Heppner reinforces: materials that are not privileged when they are created are not retroactively alchemizedinto privileged ones by later handing them to a lawyer. The protection has to attach at the moment of creation, through the attorney's involvement and a confidential channel — not be bolted on afterward.

One more bound on this: Heppner is a single district-court ruling and an issue of first impression. An SDNY court held this on these facts. It is persuasive and widely discussed, but it is not binding precedent nationwide, and the law here is still developing. Treat it as a clear signal about how a thoughtful court reasons through an AI discovery privilege ruling — not as settled doctrine.

A checklist for solos using AI in discovery

If you want the upside of AI without recreating Heppner's problem, the questions to ask before you run a matter through any tool are concrete:

• Is it attorney-directed? A lawyer should be steering the work and committing every substantive call, so the AI is assisting counsel rather than replacing the lawyer-client channel.
• Is it no-training / zero-retention? The provider terms should bar training on your data and, where available, run under zero-retention API terms — so the confidentiality the privilege needs actually exists.
• Is it audit-trailed? You want a record of who did what and when — the kind of trail you could put in front of a court to show the process was directed and controlled.
• Can you export a defensibility record? If your process is ever challenged, you should be able to hand over a tamper-evident chain-of-custody record, not reconstruct events from memory.

How Brevio is built for this posture

Brevio was designed around exactly this distinction — the issue Heppner flags is consumer-tier use without counsel direction, not any particular model. Brevio runs on enterprise AI provider APIs (OpenAI and Anthropic) under your firm's own configured keys, and, where available, zero-retention API terms; your matters are never used to train Brevio's models or any third-party model (see our Privacy Policy and the sub-processor list on our Security page). The AI suggests and you commit every call, so a lawyer stays in the loop by design — and every production can ship with a tamper-evident, SHA-256 chain-of-custody record you can hand to opposing counsel. That is the attorney-directed, audit-trailed posture Heppner rewards, built in rather than bolted on. See how the defensibility report works or start a free trial.