Does Using ChatGPT Waive Attorney Client Privilege
A federal court ordered a defendant to hand over 31 documents he made with a consumer AI chatbot. Here is what the Heppner ruling means for privilege, and the policy that prevents waiver.

If a client pastes privileged facts into a consumer AI chatbot, that act alone can waive attorney-client privilege over those facts, and anything the chatbot produces will almost certainly be discoverable. That is no longer a theoretical risk: in February 2026, a federal judge in Manhattan ordered a criminal defendant to hand prosecutors 31 documents he had created with Claude, rejecting his claim that they were privileged. The good news is that the same opinion, read together with the vendors' own contract terms, maps out exactly how to use AI on legal matters without losing privilege. This post walks through the ruling, the contract terms that decide the confidentiality question, and a written policy your firm or legal team can adopt this week.
What the court actually decided in United States v. Heppner
Bradley Heppner, a financial services executive, was indicted in the Southern District of New York in late 2025 on securities fraud, wire fraud, and related charges involving more than $150 million in investor funds. When FBI agents searched his home in November 2025, they seized devices holding what his counsel described as roughly 31 documents memorializing his exchanges with Claude, Anthropic's AI assistant. Heppner had created them on his own initiative after receiving a grand jury subpoena, and they were not thin prompts. According to his own lawyers, he "prepared reports that outlined defense strategy, that outlined what he might argue with respect to the facts and the law."
His attorneys listed the documents on a privilege log. The government moved for a ruling that they were protected by neither attorney-client privilege nor the work product doctrine. On February 10, 2026, Judge Jed Rakoff ruled from the bench for the government, and on February 17 he issued a written memorandum explaining the reasoning. He called it a question of first impression nationwide: whether a user's communications with a publicly available AI platform, made in connection with a pending criminal investigation, are protected. His answer was no.
One update since the ruling: the underlying case is no longer pending. In May 2026, a jury convicted Heppner on all counts after a three-week trial before Judge Rakoff, with sentencing scheduled for late 2026. The privilege fight over the AI documents happened pretrial; the conviction is what makes the waiver consequence concrete rather than hypothetical.
The case matters beyond one defendant. As BakerHostetler's analysis puts it, the decision confirms that longstanding privilege principles apply with full force to AI: communications shared with third-party platforms may not be confidential, and materials created outside counsel's direction fall outside work product protection. The memorandum also notes, citing industry estimates, that one major AI platform now serves more than 800 million people a week. A large share of legal AI use is happening on exactly the kind of consumer accounts Heppner used.
The three elements, and where a chatbot prompt fails
Attorney-client privilege protects communications that meet three tests: they are between a client and an attorney, intended to be and in fact kept confidential, and made for the purpose of obtaining or providing legal advice. Judge Rakoff found Heppner's AI documents failed at least two of the three, and probably all of them.
Element one: it is not a communication with an attorney
Claude is not a lawyer, and no setup changes that. The court quoted the principle that every recognized privilege presupposes "a trusting human relationship" with "a licensed professional who owes fiduciary duties and is subject to discipline." No such relationship exists, or could exist, between a user and a software platform. That defect alone disposed of the privilege claim.
Element two: it is not confidential
This is where the vendor's terms of service become evidence. Heppner used the consumer version of Claude, whose privacy policy told users that Anthropic collects their inputs and outputs, uses them to train the model, and reserves the right to disclose the data to third parties, including government authorities, even without a subpoena. Under those terms, the court held, no reasonable expectation of confidentiality could exist. It made no difference that Heppner intended the exchanges as private preparation for talking to his lawyers. He had already shared the equivalent of his notes with a third party.
The court added a second, sharper point in a footnote that every lawyer should read twice: to the extent Heppner typed information into Claude that he had learned from his attorneys, he waived privilege over that information by disclosing it to Anthropic, just as if he had shared it with any other third party. Consumer AI accounts are not a neutral storage location. They are a disclosure.
Element three: it was not for the purpose of legal advice
When the government asked Claude itself whether it could give legal advice, it answered: "I'm not a lawyer and can't provide formal legal advice or recommendations." The court treated that disclaimer as fatal to the third element. And privilege cannot be created retroactively: quoting settled Second Circuit law, the memorandum holds that non-privileged communications are not "alchemically changed into privileged ones" by being forwarded to a lawyer afterward. Saving the chat output and emailing it to your attorney does not repair anything.
Confidentiality is a contract question, so read the contract
The single most useful sentence in the Heppner memorandum for practitioners is a piece of dictum. The outcome, the court suggested, might differ where the AI tool, "unlike Claude, does not train its model on user inputs or expressly reserve the right to retain or use its user's inputs." In other words, the tier of the product you use is not an IT detail. It is the confidentiality fact the privilege analysis turns on.
Here is how the tiers actually differ, from the vendors' own documents:
- Consumer tiers (free or individual paid accounts). OpenAI states plainly that it trains its models on "data from versions of ChatGPT and other services for individuals." Anthropic's consumer privacy policy, as quoted by the court, permits training on inputs and outputs and disclosure to third parties. Treat anything typed here as potentially public. There is no privilege analysis that survives these terms.
- Business and enterprise tiers. OpenAI's enterprise privacy commitments state that by default it does not train on business data from ChatGPT Business, ChatGPT Enterprise, or the API platform, and that customers own their inputs and outputs, with AES-256 encryption at rest, TLS 1.2+ in transit, and a completed SOC 2 audit. Anthropic's Commercial Terms go further in one sentence: "Anthropic may not train models on Customer Content from Services," with a Data Processing Addendum incorporated by reference and customer content defined as the customer's confidential information.
- API access. OpenAI's terms confirm API data has not been used for training by default since March 2023, and human review is limited to abuse investigation by bound personnel. Anthropic's API sits under the same commercial no-training commitment.
- Private deployment of open-weight models. Running a model such as Llama, Qwen, or Mistral inside your own cloud or data center removes the vendor from the disclosure chain entirely. This is the strongest confidentiality posture, and it is the one banks and insurers typically choose; it is also the most expensive to operate well. Our enterprise LLM integration guide walks through what that deployment actually involves.
One warning, and it is the point most vendor marketing gets wrong. An enterprise contract repairs the confidentiality element; it does not create privilege. As Spellbook's analysis of the ruling puts it, contractual data protection binds the vendor, while privilege binds the court. A no-training clause means your prompt was not disclosed in the way Heppner's was. It does not make the chatbot your lawyer, and it does not by itself satisfy the first or third elements. What it does is keep the work product doctrine and the agency argument alive, which brings us to the doctrine that will matter more in practice.
Work product is the bigger risk, and counsel direction is the whole game
Most in-house and law firm AI use is not really about attorney-client communications; it is about analysis prepared for litigation or transactions. That makes the work product doctrine, which shelters materials "prepared by or at the behest of counsel" in anticipation of litigation, the protection that does the daily work.
Heppner lost here too, on two independent grounds: his documents were not prepared at his lawyers' direction (his counsel conceded they "did not direct Heppner to run Claude searches"), and they did not reflect counsel's mental impressions or strategy. The court was unimpressed that the documents later "affected" counsel's strategy; they did not embody it when created.
But the dictum points the other way: "Had counsel directed Heppner to use Claude, Claude might arguably be said to have functioned in a manner akin to a highly trained professional who may act as a lawyer's agent within the protection of the attorney-client privilege." That is the Kovel principle, the doctrine that already covers accountants and translators working at a lawyer's direction, extended in passing to AI. It is untested, and Judge Rakoff flagged it as argument, not holding. But it is the only route any court has sketched for AI-assisted work to sit comfortably inside privilege.
A decision issued the very same day shows the doctrine is not settled. In Warner v. Gilbarco, Magistrate Judge Anthony Patti of the Eastern District of Michigan refused to compel a self-represented plaintiff, Sohyon Warner, to produce materials she had prepared with ChatGPT for her employment discrimination suit, finding work product protection applied. The court reasoned that generative AI programs are "tools, not persons," and that uploading information to a platform is not a disclosure to an adversary of the kind that waives work product. As the White & Case comparison of the two rulings observes, what separated the outcomes was not which chatbot was used but how the party engaged with it and the procedural posture. Expect more conflict before an appeals court settles it.
Two practical consequences follow. First, if AI-assisted analysis is not directed and documented by counsel, assume it is discoverable. Second, if a waiver does happen, the cost lands in privilege review: someone has to find every place the disclosed material propagated. We have written separately about what AI-assisted privilege review costs and where it fails; prevention is far cheaper than that workflow.
A firm AI policy that survives Heppner, step by step
This is a procedure a managing partner or general counsel can adopt without new software beyond the right subscriptions.
- Classify matters before anyone prompts. Two buckets are enough: anything involving client confidences, litigation, or privileged strategy goes in the restricted bucket; public-source research, marketing, and internal administration go in the open bucket. The restricted bucket is where every rule below applies.
- Ban consumer tiers for restricted work. No client facts, documents, or strategy in personal ChatGPT, Claude, or Gemini accounts, full stop. The Heppner waiver footnote makes this the highest-risk behavior in the firm.
- Put restricted work on commercial terms. A business or enterprise subscription with a no-training clause and a signed DPA. Verify the terms yourself against the vendor's current documents, because they change; the commitments cited above are the vendors' own pages, and the links in this post go straight to them.
- Have counsel direct and document the work. The attorney defines the task, instructs the AI use, and records that instruction in the matter file. One line in a matter opening memo ("counsel directs the use of [tool] under firm enterprise terms for analysis in anticipation of [matter]") captures the agency fact the Heppner dictum says might matter.
- Keep counsel's imprint on the output. Work product protects materials that reflect the lawyer's mental impressions. Edit, annotate, and integrate AI output into attorney work product rather than archiving raw chat logs as standalone strategy documents. Raw logs of a client talking to a chatbot unsupervised are the Heppner fact pattern.
- Treat AI accounts as in scope for litigation holds. A hold reaches AI workspaces the same way it reaches email. Your e-discovery process needs to know the accounts exist and who can export them.
- Train once a year, attest in writing. The ABA's Formal Opinion 512 frames competent AI use as an ethics obligation covering confidentiality, supervision of staff and vendors, and candor. A one-page policy plus an annual attestation is evidence you took that duty seriously.
For teams that want this implemented rather than drafted, this configuration work, including enterprise account setup, matter-level routing, and audit logging, is exactly what our Claude for Legal engagements cover. Organizations with regulatory constraints that rule out any vendor cloud can go further with a private deployment on open-weight models, which keeps the disclosure chain inside your own infrastructure.
Common mistakes that create waiver risk
- Forwarding the chat to a lawyer afterward. Retroactive privilege does not exist. The disclosure happened at the prompt.
- Assuming a paid consumer plan is a business plan. Individual paid tiers can still carry training and disclosure rights. The protection comes from the commercial terms and DPA, not from a monthly fee.
- Letting clients use chatbots to "prepare" for meetings with counsel. That is precisely what Heppner did, and the reports he prepared are now in the government's hands. If a client wants to organize facts, give them a template in a privileged channel instead.
- Trusting the AI's own legal claims. Claude's disclaimer that it cannot give legal advice was literally quoted against its user in court. Do not let anyone in the firm treat chatbot output as legal advice, for competence reasons as much as privilege ones.
- Forgetting the waiver is the client's. A client who independently pastes the firm's advice into a consumer chatbot can waive privilege over it. Engagement letters should say so.
Where this reasoning does not apply
Heppner is one district court decision, binding nowhere else, and Warner v. Gilbarco went the other way on work product the same day. The counsel-direction dictum is untested, and no appellate court has weighed in. Outside the United States, common-law jurisdictions apply similar privilege principles, but the analysis will run through local evidence and data protection law, and this post is not legal advice for any specific matter. It is also worth saying what the ruling does not mean: it does not prohibit AI in legal work, and nothing in it criticizes the technology. Judge Rakoff's opinion applies old rules to new facts and reads as an invitation to use AI under counsel's direction on confidential terms, not as a warning to avoid it.
Frequently asked questions
Does using ChatGPT automatically waive attorney-client privilege? No. Privilege analysis turns on what was disclosed, to whom, and under whose direction. Using an AI tool on non-confidential material waives nothing. Typing privileged facts into a consumer account whose terms permit training and disclosure is a third-party disclosure that can waive privilege over those facts, which is what the Heppner court found.
Are ChatGPT Enterprise or Claude commercial plans privileged? No plan creates privilege, because a chatbot is not an attorney. What enterprise and commercial tiers do is restore the confidentiality element: both vendors contractually commit not to train on business-tier data. That keeps privilege and work product arguments available; it does not guarantee them.
If my lawyer tells me to use an AI tool, is the output protected? Possibly. Judge Rakoff suggested in dictum that counsel-directed AI use might function like a lawyer's agent under the Kovel doctrine, which already protects accountants and translators working at a lawyer's direction. No court has yet applied that reasoning, so treat it as the best available posture, not a settled safe harbor.
Can I make old chatbot conversations privileged by sending them to my attorney? No. The Heppner court applied settled law that non-privileged material does not become privileged by being transferred to counsel. The disclosure occurred when the information was typed into the platform.
Does deleting my chat history fix the problem? No. Deletion does not undo the disclosure that occurred at the time of the prompt, and deleting material after a litigation hold or investigation is foreseeable creates spoliation risk, which is far worse than the waiver problem.
What about AI tools built specifically for lawyers? Legal-specific tools usually run on the same foundation models under commercial or API terms, so the analysis is the same: check the underlying data terms, keep counsel in the loop, and document direction. A legal interface does not change the privilege elements. On the separate question of whether the output is accurate enough to rely on, see our review of what the contract review benchmarks actually show.
Where to go from here
The firms that get this right will not be the ones that ban AI or the ones that ignore the issue; they will be the ones with a two-bucket policy, commercial terms on restricted work, and counsel direction documented in the file. If you want help putting that in place, our Claude for Legal practice implements Anthropic's legal tooling under the commercial terms this analysis depends on, with the policy, routing, and logging configured to match.
About AI Agents Plus Editorial
AI automation expert and thought leader in business transformation through artificial intelligence.



