Are AI chats anonymized? What six AI companies strip, and what they keep

The Morse teamPublished November 4, 20268 min read

No, AI chats are not anonymized before people review them or models train on them, according to the six AI companies whose own pages we read in November 2026. OpenAI, Anthropic, Google, Meta, xAI, and Microsoft each say they remove or detach some identifiers, such as your account, email address, or phone number. The words you typed stay. Under the GDPR's own definitions, a chat with the account label removed but the text intact is pseudonymized at best, and pseudonymized data that can be tied back to someone still counts as personal data, not anonymous data.

On November 4, 2026, we read each company's privacy policy and help pages, Anthropic's research post on Clio, the GDPR's text on EUR-Lex, the US health department's de-identification guidance, an FTC staff post on "anonymous" claims, and a 2019 Nature Communications study on re-identification. Two companies earn credit. Google says exactly what it detaches, who sees the result, and how long it's kept. Microsoft lists the identifiers it removes, by name.

What's the difference between anonymized and de-identified data?

Anonymized data can't be tied back to anyone. De-identified data has had identifiers removed, and it may or may not still point to someone. The GDPR draws the line in Recital 26: data that has been pseudonymized and "could be attributed to a natural person by the use of additional information should be considered to be information on an identifiable natural person." Only data "rendered anonymous in such a manner that the data subject is not or no longer identifiable" falls outside the law. Article 4(5) defines pseudonymization as processing data so it can no longer be attributed to a person "without the use of additional information," kept separately.

US health law sets a lower bar, since the Department of Health and Human Services says data de-identified under HIPAA "retains some risk of identification. Although the risk is very small, it is not zero" (HHS), and we apply its list of 18 identifiers to therapy notes in private AI for therapists.

The FTC sets the strictest test. In a July 24, 2024 post, its Office of Technology wrote that "data is only anonymous when it can never be associated back to a person," and that FTC staff will "take action when the privacy claims they make are deceptive" (FTC).

The EU's top court added a wrinkle in September 2025. In a case under the data protection rules for EU bodies, it held that pseudonymized data isn't personal data "in all cases and for every person," since a recipient without the key may be unable to identify anyone. It also said personal opinions are "necessarily closely linked" to the person who wrote them, and that, for its duty to tell people where their data goes, identifiability is judged when the data is collected, from the point of view of the organization that collected it (Court of Justice). An AI company that received your chat under your account is that organization.

Does ChatGPT anonymize your data?

No. OpenAI says it reduces personal information. Its help center says: "Before using eligible user content to improve our models, we take steps to reduce the amount of personal information in training datasets" (OpenAI). Its privacy page says it works "to identify and remove personal identifiers from training sets wherever possible" (OpenAI). Neither page lists which identifiers, and "wherever possible" leaves room for the ones it can't find.

Three other details, read November 4, 2026:

• Reviewers can read your content. A limited number of "authorized OpenAI personnel, as well as trusted service providers," may access user content "to improve model performance (unless you have opted out)" (OpenAI). The same page asks: "Please do not enter sensitive information that you would not want reviewed or used."
• Feedback sends the whole chat. If you rate a reply, "the entire conversation associated with that feedback may be used to train our models," even after you opt out.
• De-identified copies outlive deletion. OpenAI's privacy policy removes deleted data within 30 days unless "it has already been de-identified and disassociated from your account." The same policy says OpenAI will "not attempt to reidentify the information, unless required by law."

What does each AI company remove before review or training?

Mostly the link between the chat and your account. Microsoft is the only one of the six that lists the identifiers it removes. The table covers personal accounts, read on each company's own pages on November 4, 2026.

None of the six says the chats its reviewers read or its models train on are anonymized. The words are reduce, de-link, disconnect, de-identify, and, at xAI, pseudoanonymize. xAI's sentence claims pseudonymized data "no longer identifies you," the opposite of what GDPR Recital 26 says about pseudonymized data. Microsoft's page covers only the older Copilot app. The updated app, out since August 18, 2026, points to Microsoft's privacy statement (last updated September 2026), which says Microsoft uses data, "often de-identified," to "train and fine-tune AI models."

Can de-identified AI chats be traced back to you?

Yes, sometimes, because the text does the identifying. Removing an account ID leaves every name, street, employer, diagnosis, and date you typed. HHS's list of identifiers to remove covers those of "relatives, employers, or household members," the people a chat about your life tends to name.

A 2019 study in Nature Communications by Luc Rocher, Julien Hendrickx, and Yves-Alexandre de Montjoye found that 15 demographic attributes "would render 99.98% of people in Massachusetts unique," and a long chat can hold more than 15 facts about you.

Anthropic admits gaps in its own tool. Its Clio post (December 12, 2024) says Claude is told to omit private details and that analysts see only clusters above a minimum number of users, then adds that "our systems might not catch certain kinds of private information." Its Trust and Safety team can use Clio "to identify individual accounts for further review."

De-identification can also be undone on purpose. Anthropic's privacy policy (effective September 10, 2026) says it removes your user ID from flagged chats to train its safety classifiers, but "may re-identify the Inputs or Outputs to enforce our Terms of Service or Usage Policy."

Why don't de-identified chats get deleted with your history?

Because the company can no longer find them under your name, as Google explains: "past chats already reviewed by service providers aren't deleted because they aren't connected to your Google Account. Instead, they're retained for up to 3 years" (Google). OpenAI makes the same exception for de-identified chats, and our page on who reviews AI chats covers where reviewed copies go.

How do you keep your identity out of an AI chat?

Edit what you type, and keep the copies out of training:

• Strip the text yourself. Leave out names, addresses, employers, exact dates, and account numbers, yours and other people's.
• Turn off training before the chat. Each app's setting is in our table of which AI companies train on your data.
• Don't rate replies on anything sensitive. OpenAI may train on the whole rated chat, even after you opt out.
• Take your account out of the request. Our guide to using ChatGPT anonymously compares the ways.

How Morse takes your name off before the message leaves

Every company in the table de-identifies after the fact. It receives your chat under your account, stores it, and later strips labels from a copy, while the text stays as you wrote it. Morse removes who you are before a model provider gets the message, and keeps no copy of the chat on its servers.

• Anonymous, so GPT, Claude, or Gemini gets your question without your account attached. Our proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider sees the message without knowing it's you, and may keep it for a limited time to check for abuse, under its own policy. Like every method on this page, it removes the sender, not the words, so leave names and details you'd keep to yourself out of the message.
• E2EE, for anything no one should read, de-identified or not. Your browser checks the secure enclave's hardware attestation, then encrypts the message so only that enclave can read it, and our servers relay ciphertext 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 runs open-weight models like GLM, Kimi, and DeepSeek, without web search or image generation.
• Private, the default, for chats you'd rather no host keep at all. Open-weight models like DeepSeek, Kimi, and GLM, plus xAI's Grok, get everything Anonymous mode strips, plus zero data retention: Morse has zero data retention agreements with its inference, search, and other providers. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't. Otherwise Morse relies on the provider's commitment.

There's no training pipeline at Morse for a de-identified copy to sit in, because Morse doesn't train AI on your content, sell your data, or show ads. Our servers don't keep your messages or log what you write. Your chat history is stored only in your browser, encrypted. Clearing your browser or deleting your account deletes your chats, and Morse can't recover them, so keep anything you need. Because we don't store chats, we usually have very little to give authorities, as our privacy policy says.

Send a first message from the Morse home page without an account, or sign up free. E2EE chats come with Pro, at $19 a month.