If I or my staff use ChatGPT, Claude or another AI chatbot to think through a legal dispute, are those chats protected by attorney-client privilege or work product, or can the other side get them in discovery and use them as evidence?

What should I check, step by step?

1. Treat a consumer chatbot like email to an outside party: don't paste privileged advice, settlement positions or your lawyer's strategy into it.

2. If AI is used for litigation work, have counsel direct it and record that direction. Heppner left open that AI used at a lawyer's direction might act like the lawyer's agent; it turned on the defendant acting on his own.
3. Prefer enterprise or API plans whose terms bar training on your inputs, restrict disclosure and let you delete data. Morgan's protective order allowed uploading confidential material only to providers bound by those three terms.
4. Read the provider's privacy policy and your data-control settings. Heppner relied on the consumer policy's permission to collect, train on and disclose chats; the opinion didn't address whether turning off training would have changed that.
5. Add AI tools, consumer and enterprise, to litigation hold notices once a dispute is likely. Chats you can log in and export are likely within your possession, custody or control, and deleting them after the duty to preserve attaches invites sanctions under Rule 37(e).
6. Don't rely on deleting a chat to make it disappear. Providers keep logs for their own purposes and can be subpoenaed or ordered to preserve them.
7. If you withhold AI material as work product, expect to log it document by document. Fontainebleau rejected a blanket objection.
8. If an AI answer could matter later (advice to a customer, a figure in a report, an instruction it gave an agent), save the exact output and the model version when it happens, rather than hoping to recover it from the provider.
9. Ask a litigator about your matter. Three of these four decisions come from magistrate or trial judges, they disagree, and none binds courts elsewhere.

Are ChatGPT or Claude chats protected by attorney-client privilege or work product?

Heppner: after a grand jury subpoena, a securities-fraud defendant used consumer Claude on his own to prepare reports on his defense, then shared them with his lawyers; the FBI seized them in a search of his home. The court ruled from the bench on February 10, 2026 and explained in a February 17 opinion that the documents lacked at least two of the privilege's three elements: they weren't communications with a lawyer, they weren't confidential, and they weren't made to get legal advice from counsel. Work product failed because they weren't prepared by or at the behest of counsel and didn't reflect counsel's strategy.

• Warner v. Gilbarco: a self-represented employment plaintiff used ChatGPT to prepare her case. The court refused to compel production of her AI use, treating it as material prepared in anticipation of litigation, and held that work-product protection is waived only by disclosure to an adversary or in a way likely to reach one. It said ChatGPT and other generative AI programs 'are tools, not persons'. Morgan v. V2X followed Warner for another self-represented plaintiff, but ordered him to identify the AI tool he used and limited uploads of confidential discovery material to providers contractually barred from training on inputs, restricted on disclosure and offering deletion.
• Fontainebleau Florida Hotel v. Botach: in a defamation case, the plaintiff sought the defendant's AI communications and proposed subpoenas to AI providers. On October 21, 2026 the court held that work product can cover AI prompts and outputs even without a lawyer's direction, declining to follow Heppner as a categorical limit, but overruled blanket objections to several requests, required a privilege log for anything withheld, and permitted Rule 45 subpoenas to four providers limited to communications since December 3, 2024 on set topics, after the parties confer on how providers produce them so protection can still be claimed.

Can OpenAI or another AI company be forced to hand over your chat logs?

In the consolidated copyright cases against OpenAI in New York, the court noted OpenAI keeps tens of billions of ChatGPT logs in the ordinary course of business. On January 5, 2026 the district judge affirmed orders requiring OpenAI to produce a 20 million log sample, de-identified and under a protective order, rejecting OpenAI's proposal to search it and produce only matching conversations. The Heppner opinion cited an earlier ruling in that case for the point that users have limited privacy interests in conversations they voluntarily disclose to a platform that retains them.

• A provider's own copy is outside your control and its retention isn't designed for your dispute. If an AI answer becomes contested, the version the provider keeps may be gone, de-identified or not produced, and the version you screenshot is easy to question.

What should AI app builders do about user chat logs being discoverable?

Your users' conversations are records you hold. Decide and publish how long you keep prompts and outputs, include AI logs in your litigation-hold process, and know who can export them, because a court can treat what your users or staff can retrieve as within their control.

• When an AI answer is relied on, such as a support bot's statement to a customer or an agent's instruction, record at the time which model and version produced which output. A hash-based record lets you show later that a saved transcript is unchanged without handing the text to yet another third party.