Heppner and the Press: Are a Journalist's AI Chats Protected?
Shield laws protect reporters, not the AI company holding their chats. Prosecutors have already used a warrant to get one defendant's ChatGPT records.
What Heppner decided
In United States v. Heppner, a federal judge in New York faced a narrow question. A fraud defendant had used the consumer version of Claude, Anthropic's chatbot, to work on his own defense. Agents executing a search warrant seized his devices, which held the AI documents and his prompts.[2] He argued they were covered by attorney client privilege, the rule that keeps talks between a client and a lawyer secret so clients can speak freely.
Judge Jed Rakoff disagreed in a written opinion on February 17, 2026.[1] Claude is not a lawyer, its terms disclaim giving legal advice, and its privacy policy lets the company share chats with third parties, including the government.[3] (I argued that ruling was a mistake in Heppner Was a Bad Call. This post asks a different question.)
Two different protections
Reporter's privilege is not attorney client privilege. One protects a private relationship. The other protects a journalist from being forced to name a source. It comes from state shield laws and a limited federal doctrine, not from confidentiality between two people.[4]
The Supreme Court set the limits in Branzburg v. Hayes in 1972. Reporters have no absolute right to refuse to testify, though many courts apply a qualified privilege that the government can overcome.[5]
Can the AI company be made to hand it over?
This is the part Heppner does not answer. That ruling was about files taken from the defendant, not a demand served on the company. A demand to the company is different. It asks a third party for its records, and a shield law protects the journalist, not the vendor.[4]
Here a different law matters. The Stored Communications Act generally bars service providers from handing over the content of stored messages, and courts apply that bar to civil subpoenas too.[6] So a private party suing a reporter usually has to ask the reporter, where a shield law can push back. The government is different. It can get prompts and responses from an AI company with a warrant.[7]
That has already happened. In June 2026, a federal judge in Manhattan refused to block a search warrant to OpenAI for a fraud defendant's ChatGPT records in United States v. Kim. He can still fight over specific items later.[8] A reporter is not a defendant, but the path is the same: the records sit with the company, and the company can be served.
The missing federal shield
There is still no federal shield law. The PRESS Act would create one. The House passed it unanimously in January 2024,[9] then a single senator blocked it in the Senate.[10] It would have covered more than reporters, barring the government from forcing telecommunication providers to give up a journalist's sources and work product too.[11] As of November 2026, press freedom groups are still asking Congress to pass it.[12]
That leaves reporters guessing. Sharing notes with an AI is ordinary now. A reporter who pastes a source's name into a chatbot may have handed it to a company that can be served with a warrant. That is bad for the public, because sources dry up when a reporter cannot promise protection.
But do not overstate the law. I found no case that decides a journalist's AI chats directly. Shield laws differ by state, and the PRESS Act is not law. Until that changes, the safe rule for a reporter is blunt. Assume the chatbot is a witness.
Sources
- Debevoise & Plimpton, Feb. 18, 2026, "Update: Judge Rakoff Issues Written Opinion That AI Generated Documents Are Not Protected by Privilege" (from the DeepSeek draft; not opened here)
- Ogletree Deakins, 2026 (agents executing a search warrant seized Heppner's devices holding the AI documents and prompt logs)
- Duane Morris, March 2026 (Anthropic's terms disclaim legal advice; privacy policy allows disclosure to third parties including government)
- Wikipedia, "Shield law" (from the DeepSeek draft)
- Wikipedia, "Branzburg v. Hayes" (1972)
- Optiver Australia v. Tibra Trading, order granting in part motion to quash (N.D. Cal. Jan. 2013): civil subpoenas are subject to the Stored Communications Act's bar on providers disclosing content
- Stanford Center for Internet and Society, March 2026, "Applying Federal Digital Communications Privacy Law to Hosted AI Models" (the SCA requires a warrant for prompts and responses; OpenAI's "get a warrant" position)
- Reuters via Sahm Capital, June 23, 2026, "Judge allows US search warrant targeting executive's AI chatbot records" (United States v. Kim, No. 1:25 cr 00359, S.D.N.Y., Judge Schofield)
- Rep. Kevin Kiley, Jan. 18, 2024: the House passes H.R. 4250, the PRESS Act
- Voice of America, 2024, "Bill to protect journalists fails in US Senate" (blocked by Sen. Tom Cotton; no federal shield law in place)
- NationofChange, Jan. 24, 2024 (the PRESS Act would bar the government from compelling journalists and telecommunication providers to disclose sources and work product)
- Freedom of the Press Foundation, Nov. 2026, "New Congress must prioritize press freedom" (from the DeepSeek draft; not opened here)
Inspired by Happy. Prepared by DeepSeek (version unconfirmed). Reviewed and edited by Claude.
Next relatedHeppner Was a Bad Call: A Chat You Mark Privileged Should Stay PrivateRevision history (R0 current)
| Rev | Date | What changed | Page |
|---|---|---|---|
| D1 | Nov 19, 2026, 8:58:34 pm CST (received) | What DeepSeek returned, unedited (self report: deepseek-flash, version UNSURE, search on). Prompt handed out at 8:52:16 pm CST, so about six minutes to draft. | view |
| R0 | Nov 19, 2026, 8:59:37 pm CST (reviewed and posted) | First publish, reviewed and edited by Claude (Opus). Fixed the seizure wording and the Stored Communications Act explanation; added United States v. Kim (warrant to OpenAI) and the PRESS Act's 2024 House vote and Senate block; dropped an unconfirmed Claude Wikipedia link and the Harvard Law Review source. Sources checked through search results. | view |