AI Agents for Lawyers

Regulation & Ethics

Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots

Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. What Judge Rakoff decided in United States v. Heppner
  2. Warner v. Gilbarco: the same question, the opposite answer
  3. The 20 million ChatGPT logs: deletion is a setting, retention is the law
  4. "In the public domain": the UK and continental view
  5. Where privilege stands, scenario by scenario
  6. Seven rules for AI use that keeps privilege intact
  7. The "AI privilege" idea and why it will not rescue anyone
  8. What to do this week

Attorney–client privilege does not attach to what a client or lawyer types into a consumer chatbot. In United States v. Heppner (S.D.N.Y., 17 February 2026) Judge Rakoff held that a defendant's Claude memos were never privileged, because Claude is not an attorney and its terms permit logging and disclosure. Privilege can survive when a lawyer directs the tool, under enterprise terms, inside the engagement.

What Judge Rakoff decided in United States v. Heppner

The facts are the kind every litigator will recognise. A defendant facing federal charges in the Southern District of New York took the advice his lawyers had given him, went to a consumer Claude account and used it to draft memos on his defence strategy. The FBI seized 31 such documents. His counsel claimed privilege and work-product protection over all of them. Judge Rakoff rejected the claim from the bench on 10 February 2026 and set out his reasons in a written opinion on 17 February 2026; the Harvard Law Review Blog has the fullest account.

Three holdings matter.

First, the memos were not communications with a lawyer. In the court's words, "Claude is not an attorney". Attorney-client privilege protects confidential communications between client and counsel for the purpose of obtaining legal advice; a conversation with a language model is neither.

Second, there was no reasonable expectation of confidentiality. Anthropic's consumer terms permit logging and disclosure of user content, and since 28 August 2025 consumer users have had to choose whether their chats are used for training, with opted-in data kept for five years. A client who types privileged advice into a service on those terms has disclosed it to a third party who is free to keep it and, in some circumstances, to hand it over.

Third, the documents were not work product, because they were not prepared by or at the direction of counsel in anticipation of litigation. They were the client's own research.

The precise framing matters. A Fortune op-ed of 16 May 2026 described a Manhattan court finding that a defendant had "waived" privilege by preparing his case with a chatbot (Fortune). The opinion is stricter than that: privilege never attached, so there was nothing to waive. That closes off the usual arguments about inadvertent disclosure and clawback. The one door Rakoff left open is the Kovel route, named after United States v. Kovel (2d Cir. 1961), under which a non-lawyer retained by counsel to assist in giving legal advice falls inside the privilege. Had the lawyers directed the AI use as part of their own work, the analysis might have been different.

The Harvard Law Review's commentator objects that privilege should not turn on a vendor's terms of service that consumers never read. The objection is fair and, as of September 2026, beside the point for practice: the terms decide the outcome until an appellate court says otherwise.

Warner v. Gilbarco: the same question, the opposite answer

Within about a week, a court in the Eastern District of Michigan reportedly reached the opposite result in Warner v. Gilbarco (E.D. Mich., February 2026). The decision has so far been reported by a single vendor blog, so treat the details as reported rather than verified. On that account, the court described ChatGPT and other generative AI programs as "tools, not persons", and held that work product prepared with AI assistance stays protected, on the reasoning that using a tool is no more a disclosure than using a word processor.

The two decisions are not in conflict once you look at who was using the tool and on what terms. In Heppner the client used a consumer service alone, outside the engagement. In Warner, as reported, lawyers used AI inside their own work product. The variables that decided both cases were direction by counsel and the confidentiality of the channel. That is also the structure of the Kovel doctrine, and it gives lawyers a recipe rather than a warning.

The 20 million ChatGPT logs: deletion is a setting, retention is the law

The second reason to take the channel seriously is In re OpenAI Copyright Litigation, the multidistrict proceeding in the Southern District of New York that consolidates 16 suits, including the New York Times's. In November 2025 Magistrate Judge Wang ordered OpenAI to produce a sample of 20 million de-identified ChatGPT conversation logs, and on 5 January 2026 Judge Stein affirmed the order. The plaintiffs had asked for 120 million; the 20 million represent about 0.5% of the tens of billions of conversations OpenAI had preserved under a litigation hold (National Law Review; Lawyer Monthly). Judge Stein's reasoning was that users had "voluntarily submitted their communications" to OpenAI. The logs are discoverable electronically stored information in a third party's hands, with no privilege of their own.

Two things follow for a lawyer. A user's deletion of a chat is a product setting, and a preservation order overrides it; anything a lawyer or client typed into a consumer ChatGPT account during the hold period is sitting in a database a court has already ordered sampled. And enterprise tiers change the position only as far as their contracts go. Anthropic does not retain API content by default and offers zero data retention by arrangement, but flagged content can be kept for up to two years, and the Compliance API feed and the local session transcripts from Claude Code and Cowork are retained for six years by default; as of 18 September 2026 the Compliance API also returns Claude-in-Chrome transcripts to Enterprise administrators (reported by Releasebot). Those firm-held logs are matter records. They are privileged to the extent the underlying work is, discoverable to the extent it is not, and they belong on the same retention schedule as the file.

"In the public domain": the UK and continental view

English law asks the same question with a blunter answer. In UK v SSHD [2026] UKUT 81 (IAC) the Upper Tribunal said that pasting client letters into public AI tools puts them "in the public domain". Legal professional privilege depends on confidentiality, and confidentiality is lost the moment material enters a service that may retain, train on or disclose it. The SRA's warning notice of 17 August 2026 cites the case, and the judiciary's guidance of 31 October 2025 tells judges not to enter private information into public AI tools either. The SRA notice has its own page.

Germany and Austria approach the point through professional secrecy rather than privilege. Under § 43a Abs. 2 BRAO a Rechtsanwalt is bound to confidentiality, and the BRAK's guidance of December 2024 says a tool that stores or forwards client data for model training breaches that duty outright. Austria's § 9 RAO imposes the same obligation, and the ÖRAK's four practice rules from March 2026 (EU servers, disclose AI use, the lawyer keeps final responsibility, no client data for training) translate it into settings. The CCBE's guide of 2 October 2025 adds that clients should be informed and allowed to object. For EU competition investigations there is a further wrinkle in Akzo Nobel (C-550/07 P, 2010): in-house counsel's communications carry no privilege at all, with or without AI, so an in-house team feeding an EU-investigation file into any tool has confidentiality to protect but no privilege to lose. The Austrian and German bar guidance and the GDPR analysis cover the data-protection layer underneath.

Where privilege stands, scenario by scenario

ScenarioUnited States (as of September 2026)England and Wales, Germany, AustriaWhat decides it
Client alone types privileged advice into a consumer chatbotNever privileged; not work product (Heppner)Confidentiality lost; material "in the public domain" (UKUT 81)No lawyer, no confidential channel
Lawyer uses a personal consumer account for client workPrivilege over the underlying advice is at serious risk; the prompts themselves are discoverable ESI (OpenAI logs order)Breach of the duty of confidentiality; SRA and bar exposureTerms permit retention, training, disclosure
Lawyer uses a firm enterprise tier with a DPA, no training and defined retentionProtected, on the Warner reasoning as reported; the tool is like a word processorPermitted, subject to bar rules on EU servers and disclosureConfidential channel; counsel directs the use
Lawyer directs an agent to draft, for example, a privileged investigation memoProtected as work product prepared at counsel's direction; Kovel logic appliesProtected, provided the platform meets the secrecy rulesDirection by counsel; documented instruction
AI meeting recorder on a client callPrivilege depends on consent and control of the transcript; NYC Bar Opinion 2025-6 (22 Dec 2025) requires consent of all parties and defaults to not recordingSame confidentiality analysis; the transcript is a client recordWho holds the transcript and on what terms
Client asks a chatbot to "check" the firm's adviceNever privileged; the advice itself may now be in third-party logsConfidentiality lost for that copyClient acted outside the engagement

The meeting-notes page deals with the recorder scenario in detail; the confidentiality checklist covers the settings behind the third and fourth rows. The fourth row is not hypothetical: the employment plugin in Claude for Legal ships a workflow for privileged investigation memos, and whether its output is privileged depends entirely on the structure around it, not the label on the file.

Seven rules for AI use that keeps privilege intact

  1. Counsel directs, counsel documents. Every AI task on a matter is initiated by a lawyer and recorded as part of the lawyer's work. Under Kovel logic the tool is an assistant retained by counsel, not a confidant of the client.
  2. Enterprise channel only. Client material goes only into tiers with a signed data processing agreement, no training on inputs and a stated retention window. ABA Formal Opinion 512 (29 July 2024) requires informed client consent where a self-learning tool may expose client data (ABA, PDF); the cleaner answer is never to use such a tool for client data at all.
  3. Label the work. Outputs generated at counsel's direction carry the same "privileged and confidential, attorney work product" header as any other draft, and are stored in the matter workspace, not in a chat history.
  4. Own the logs. Know where prompts and outputs are retained, for how long and by whom; put firm-held logs on the matter retention schedule; ask the vendor for zero data retention where the models you use allow it.
  5. Tell the client what not to do. The Heppner facts began with a client acting alone. The engagement letter should say, in plain terms, that privileged advice must not be pasted into personal AI accounts.
  6. Treat the proposed California test as the standard. COPRAC's proposed comment to Rule 1.6, approved for public comment on 13 March 2026, would treat exposing client data to AI as "revealing" it where there is "a material risk of unauthorized access, retention, or use". If a channel fails that test, it fails.
  7. Record consent for recordings. Meeting assistants follow NYC Bar Opinions 2025-6 and 2026-2: consent from all parties, no surreptitious recording, and a default of not recording.

A clause for the engagement letter, written so that a client can actually follow it:

Use of AI tools. We use AI tools operated under contracts that
prohibit the provider from training on your information and that
limit how long it is retained. These tools are used at the direction
of your lawyers as part of our work for you, and their outputs are
reviewed by a lawyer before they are relied on.

To preserve legal privilege and confidentiality, please do not enter
our advice, our correspondence with you, or documents we have
prepared for you into any personal or consumer AI service (for
example a personal ChatGPT, Claude or Gemini account). A US federal
court held in February 2026 that material a client prepared in a
consumer AI service was not privileged. If you would like AI
assistance with any aspect of this matter, ask us and we will
arrange it inside the engagement.

The sixth module of the AI Lab for Lawyers course, on risks and regulation, covers the privilege and confidentiality questions this page raises.

The "AI privilege" idea and why it will not rescue anyone

Sam Altman has argued that conversations with AI should enjoy a doctor- or lawyer-style privilege. As of September 2026 no legislature has adopted anything of the kind, and commentators who have examined the proposal argue that whatever protection users get will come from vendor code and contract terms long before it comes from courts (International Center for Law & Economics; Artificial Lawyer). Even if such a privilege existed, it would belong to the user's relationship with the model, not to the lawyer-client relationship, and the OpenAI logs order shows how courts currently treat the underlying data: as ordinary ESI in a third party's hands. The working assumption for the next several years is that only the traditional privilege exists and that it protects only what a lawyer handles through a confidential channel.

What to do this week

  1. Add the engagement-letter clause above, or your own version, to new matters, and send a one-paragraph note to existing clients in live disputes.
  2. Check which tiers your lawyers are actually using; personal ChatGPT, Claude and Gemini accounts are the Heppner channel. The state bar opinions and ABA Opinion 512 give you the language for the ban.
  3. Write down where every AI log lives and for how long, and add it to the matter retention schedule.
  4. Set up matter-level workspaces so that AI work is done inside the engagement, at counsel's direction, with the Claude Projects setup or its equivalent in your platform.
  5. For litigators, decide now how you will answer a discovery request for AI prompts on a matter; the litigation practice page has the workflow, and the short answer on waiver is what to send a client who asks.

Frequently asked questions

Does using ChatGPT waive attorney-client privilege?

For a client acting alone on a consumer account, the better US view after Heppner is that privilege never attaches, so there is nothing to waive, and the material is discoverable. For a lawyer using an enterprise tool at their own direction inside the engagement, the reported Warner decision treats the tool like a word processor and protects the work product. The channel and who directs the use decide it.

What did the court decide in United States v. Heppner?

Judge Rakoff in the Southern District of New York ruled from the bench on 10 February 2026 and in writing on 17 February 2026 that memos a defendant had prepared with consumer Claude, using his lawyers' advice, were neither privileged nor work product. Claude is not an attorney, the terms allowed logging and disclosure, and counsel had not directed the work. He signalled that a lawyer-directed, Kovel-style use might be treated differently.

Are my ChatGPT conversations discoverable in litigation?

Yes, as electronically stored information. In the OpenAI copyright multidistrict litigation a magistrate judge ordered 20 million de-identified logs produced in November 2025 and Judge Stein affirmed on 5 January 2026, reasoning that users voluntarily submitted their communications to OpenAI. A user's deletion is a product setting; a preservation order overrides it.

Can a lawyer use AI on a matter without losing privilege?

Yes, if the structure is right. The lawyer initiates and documents the task, the tool runs on an enterprise tier with a data processing agreement, no training on inputs and a known retention window, outputs are stored in the matter file and reviewed by a lawyer, and the client is told in the engagement letter not to paste the firm's advice into personal AI accounts.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion