Private AI for journalists: source names, leaked files, and interview audio

The Morse teamPublished October 26, 20266 min read

Private AI for journalists means keeping three things out of other people's records: your sources' names, the documents they gave you, and your interview audio. As of October 2026, the safe default is to keep all three out of consumer AI tools, newsroom AI accounts, and cloud transcription services, because each one stores a copy that legal process or an administrator can reach, and the federal bill that would have shielded those copies died in the Senate.

We read the federal statute, the PRESS Act's text and record, the Justice Department's news media rule, and the AI and transcription companies' own pages on October 22, 2026. Where a company's site blocked us, we say which archive copy we read instead. This is privacy guidance for using AI in reporting, not legal advice and not advice on how to report.

Is it safe for journalists to use ChatGPT?

For public material, mostly yes. For anything that could identify a source, no, and press groups say the same. The Associated Press's generative AI standards, from August 2023, tell staff: "We urge staff to not put confidential or sensitive information into AI tools" (AP, read in an October 12, 2026 archive copy because ap.org blocked us). The Freedom of the Press Foundation's advice on transcription, updated June 1, 2026, says cloud tools fit content "that isn't particularly sensitive," and that for topics "too sensitive to entrust to a third party" its author runs transcription on his own device (Freedom of the Press Foundation).

The general mechanics, such as training settings, temporary chats, and human review, are the same for reporters as for anyone, and we cover them in how to use AI privately. For a journalist, the person someone may come looking for is the source who trusted you.

Which reporting tasks put a source at risk?

We matched the five ways source material reaches an AI tool to where it ends up, using the pages cited in the sections below, all read October 22, 2026.

Hiding your own identity doesn't hide the source's. The prompt still carries the source's name, and a leaked document can identify the leaker by its contents alone.

Does a shield law protect what you type into an AI?

Not clearly, and not at the federal level. Three pieces of US law apply. This isn't legal advice; your newsroom's lawyer is the person to ask about a live story.

• The Privacy Protection Act of 1980 bars government officers from searching for or seizing a journalist's work product in a criminal investigation, with exceptions, such as when the journalist is suspected of the crime or a seizure is needed to prevent death or serious injury. It protects materials "possessed by a person reasonably believed to have a purpose to disseminate to the public" a newspaper, broadcast, or similar work (42 U.S.C. 2000aa). Its ban covers searches and seizures, and it was written decades before chat history sat on someone else's server.
• The PRESS Act would have gone further. H.R. 4250 cleared the House Judiciary Committee 23 to 0 on July 19, 2023, passed the House by voice vote on January 18, 2024, and was referred to the Senate Judiciary Committee on January 22, 2024, where it stopped (Congress.gov). Its section 4 barred federal entities from compelling a "covered service provider" to hand over records "stored by a covered provider on behalf of a covered journalist," except to stop a "reasonable threat of imminent violence," and it defined a provider as anyone that "by an electronic means, stores, processes, or transmits information in order to provide a service to customers" (bill text). That wording is broad enough on its face to include an AI chat service. We found no reintroduction in the 119th Congress as of October 22, 2026.
• The Justice Department's own rule, 28 C.F.R. 50.10, revised effective May 2, 2025, requires the Attorney General's approval before prosecutors use a subpoena or court order to get a journalist's "communications records" from a "third-party communication service provider," and it usually requires notice to the journalist first. It also says it "does not, create any right or benefit" enforceable in court (eCFR).

For how AI chats have already reached courts, from warrants to Heppner, see can AI chats be subpoenaed.

What can your newsroom's admin see in an enterprise AI account?

Often every chat, since ChatGPT Business and Gemini on Google Workspace both let admins export or hold employees' conversations (OpenAI, read in an October 2, 2026 archive copy because openai.com blocked us; Google). Those controls serve the newsroom's own legal holds, but they make a source's name typed into a work chat your employer's record, kept on its schedule.

Are AI transcription tools safe for interviews?

For an on-the-record interview you'll publish in full, usually. For a confidential source, it depends on the service's terms. Otter's privacy policy, effective June 16, 2026, lists "training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)" among its uses, and names "Data labeling service providers" and "Artificial intelligence service providers" among those it shares with (Otter). Otter's Data Request Policy, effective October 1, 2022, says it discloses user content on "a legal process, such as a subpoena, court order, or search warrant," and that it "will notify Customer before disclosing" unless it is barred from doing so (Otter).

OpenAI's law enforcement policy asks for "a valid warrant or equivalent" before giving police your content (OpenAI), while Otter's names a subpoena as one route. A recorded voice is also harder to de-identify than a name in text. The Freedom of the Press Foundation uses Otter for non-sensitive audio and points reporters to on-device tools such as OpenAI's Whisper for the rest (Freedom of the Press Foundation).

How Morse keeps source material off its own servers

Every row in the table starts with a copy of your reporting on a server someone else controls, where a subpoena, a legal hold, or an admin can reach it. Morse is built not to keep that copy: your chat history is stored only in your browser, encrypted, and our servers don't keep messages or log what you write. If a demand arrives, our privacy policy says we "usually" have "very little to give" authorities.

• E2EE, for the leaked document no one should read. Paste the passages you need into a chat started in E2EE. Your browser checks the secure enclave's proof, then encrypts each message so only that enclave can open it, and our servers relay text they can't read. The enclaves are run by third-party providers, not by Morse, and the guarantee depends on the enclave hardware and attestation working as designed. E2EE comes with Pro, $19 a month.
• Private, for background research on a story. It's the default for a new chat, and Morse has zero data retention agreements with its inference, search, and other providers.
• Anonymous, to ask GPT, Claude, or Gemini without the provider knowing it's you. The provider still reads the text and may keep it for a limited time to check for abuse, under its own policy, so describe a source by role, never by name.

Your history is stored only in your browser, so lock your device, sign out on shared machines, and save anything you'll need, because clearing the browser deletes it and Morse can't recover it.

You can try a chat from the Morse home page without an account, or start free and move to Pro when a story needs E2EE.