AI Chats Are Discoverable: Court in Heppner Denies Privilege── What employees type into AI about a dispute can be produced in U.S. litigation, and records made outside a lawyer's direction carry no privilege

On 26 October 2026, the American legal publication Law.com reported that companies are now facing discovery requests for employees' generative AI chats, uploads, outputs and usage logs. Do the records of what employees type into generative AI become documents the other side can obtain in litigation? In U.S. litigation they can. AI exchanges can count as electronically stored information, and in the February 2026 Heppner case a court refused to treat a defendant's records with Claude as privileged.

01Records of AI conversations are documents that can be produced in U.S. litigation

The Law.com piece ran in Legaltech News, its section on law and technology. It reports that organizations are being asked to hand over AI chats, uploaded files, AI outputs and usage logs. It also notes that employees turn to AI before and after a dispute arises, to summarize what happened or to test their own arguments.

U.S. civil litigation includes a process called discovery. Each party must produce documents and electronic records relevant to the case when the other side asks for them. Exchanges with an AI system can fall within those electronic records.

The protections that usually shield sensitive records have already failed for AI records at least once. In February 2026 a federal court in New York ruled in the Heppner case. About 31 documents that the defendant had created with Claude, without direction from his lawyers, were held not to be covered by attorney-client privilege, the right to keep communications with one's lawyer from the other side.

Anyone who consults AI about a dispute at work should settle three things before typing. Which AI tool to use. How its history will be kept. When to bring in a lawyer.

02Rule 34 reaches any data compilation, stored in any medium

The basis for treating AI records as discoverable is Rule 34 of the Federal Rules of Civil Procedure. The rule sets out what a party in a civil case in federal court may be asked to produce.

Rule 34 defines its scope broadly. It covers documents and electronically stored information, including writings, drawings, graphs, charts, photographs, sound recordings and images. It also covers other data and data compilations, stored in any medium.

The chats, uploads, outputs and usage logs that Law.com lists are all data kept on servers. Nothing in the wording of Rule 34 sets AI records apart. So AI records can fall within this definition.

The scope of this essay needs stating too. Rule 34 governs civil cases in U.S. federal courts. Heppner, discussed below, was a criminal case, and I use it only as an example of how privilege was decided. I have not examined procedure in Japanese courts.

The path from an employee's AI conversation to the other side's hands can be drawn as a flow.

03Those asked for AI records range from an individual defendant to an AI provider

The wording of Rule 34 is broad. Whose AI records have actually been sought? Two cases show the range.

The first is Heppner. The defendant in a criminal case had exchanged messages with Claude, the AI system made by Anthropic, and turned the content into documents. Those documents were seized, and the treatment of about 31 of them was contested.

The second is the copyright suit brought by The New York Times against OpenAI. There, OpenAI, the provider of ChatGPT, was ordered to preserve and then produce logs of its users' conversations.

The two cases differ in who held the records and in the kind of case. Yet in both, the court treated AI records, whether an individual's or a provider's, as material that could reach the other side. Rule 34 does not pick industries. For any company facing civil litigation in U.S. federal courts, a pharmaceutical company included, employees' use of AI sits under the same rule.

04Privilege failed because the counterpart was not a lawyer and confidentiality was never promised

Why were Heppner's 31 or so documents not protected? The ruling was written by Judge Jed Rakoff of the federal district court. I have not read the opinion itself. What follows draws on the reasons set out in a commentary by the law firm Orrick.

The first reason is that Claude is not a lawyer. Attorney-client privilege protects communications between a client and an attorney. In the words Orrick quotes, the fact that Claude is not an attorney alone disposes of the privilege claim.

The second reason is that confidentiality could not be expected. Anthropic's privacy policy says it collects users' inputs and may disclose them to third parties. The judge concluded that someone typing under those terms had no reasonable expectation that the content would stay secret.

The third reason is that Heppner used Claude on his own initiative, not at his lawyers' direction. On that basis the judge also rejected work-product protection, the doctrine that shields materials prepared by or for counsel in anticipation of litigation.

The judge gave the three reasons separately. For the first, at least, he said it was enough on its own to defeat the claim. On this reasoning, what a person confides, without a lawyer's involvement, to something that is not a lawyer falls outside attorney-client privilege.

05OpenAI was ordered to keep logs slated for deletion and later to produce 20 million of them

What happens to records that privilege does not protect? In The New York Times' case against OpenAI, the steps from an order to keep records to an order to hand them over can be followed by date.

Keeping records relevant to a lawsuit instead of deleting them is called preservation. On 13 May 2025 the court ordered OpenAI to preserve output logs, including logs it had planned to delete, and to keep them segregated.

That ongoing duty ended on 26 September 2025 by agreement of the parties. From that date OpenAI no longer had to preserve new logs.

In January 2026 Judge Sidney Stein of the federal district court in New York upheld earlier orders requiring OpenAI to produce 20 million de-identified logs. The 20 million were a sample drawn from user conversations. The production came with de-identification and a protective order limiting how the logs could be used.

Even for logs held by an AI provider, the court ordered preservation and then production. Users' conversations were to go to the opposing party, with privacy safeguards attached.

06The tool, the retention setting and a lawyer's involvement decide the treatment of records

AI records get preserved and produced. What can users decide before they type? Three things.

A consumer tool or an enterprise tool

Judge Rakoff treated Anthropic's terms, which permit disclosure of inputs, as a reason that confidentiality failed. On this reasoning, with a tool whose terms allow inputs to pass to third parties, a claim of confidentiality is hard to make. Would an AI used under an enterprise contract fare better? No court has decided that yet. Orrick also advises checking with a lawyer before use.

A setting that keeps history from being deleted

Rule 37(e) addresses electronically stored information that should have been preserved but was lost. It applies when a party, in anticipation or during the conduct of litigation, failed to take reasonable steps to keep it. If AI histories are left on automatic deletion after litigation is anticipated, that rule can come into play. A commentary by the law firm Duane Morris recommends having a way to suspend deletion ready in advance.

Use at a lawyer's direction

Judge Rakoff said that if Claude had been used at counsel's direction, there might have been room for a different analysis. He did not say what the answer would be. Bringing a dispute to a lawyer before turning to AI leaves more room to argue over how the records are treated.

Which tool was used and whether a lawyer directed its use cannot be changed after the typing is done, and a setting to stop deletion comes too late once the records are gone. The tool, the retention setting and the call to counsel all have to be settled as soon as a dispute appears likely.

07Heppner is one trial-court ruling, and enterprise AI records remain an open question

Judge Rakoff suggested a lawyer's involvement might change the outcome. No court has yet tested that difference.

Heppner is described as the first ruling of its kind to deny privilege to AI conversations. It is still the decision of a single federal district court. Whether appellate courts or other district courts will reach the same result is unknown. Three questions remain open.

My own expectation is this. As companies use more AI, requests for AI records are likely to grow, and Law.com reports that they have already begun. Which courts rule, on which tools, and to what effect is what to watch next.

What employees type into generative AI can become documents the other side obtains in U.S. litigation. What shapes how those records are treated is the tool's terms, the ability to stop deletion and a lawyer's direction.

All three must be settled before typing. The next time I am tempted to have AI summarize a dispute, I intend to ask first who I should be talking to instead.