Who Can Read Your AI Chats?

A thumbs-up can override your AI privacy settings. Who can see your chats, how long they’re kept, and a 10-minute check.

Some of your AI chats can be read by people working for the company behind the app, and some copies, disconnected from your account, can be kept for years after you delete them. A button under the answers helps decide which chats.

Take the chat you’d least like anyone to read. A worry about a lump you haven’t shown a doctor, say. The figures behind a debt. A draft email about a colleague. You typed it into ChatGPT late one evening, the answer was better than you expected, and you clicked the thumbs up.

You switched training off months ago, so you assumed the chat stayed between you and the app. OpenAI’s help page asks your exact question, “What happens if I submit feedback after opting out of training?”, and answers it: “If you choose to provide feedback, such as selecting thumbs up or thumbs down on a response, the entire conversation associated with that feedback may be used to train OpenAI models” (OpenAI Help, Data controls in ChatGPT).

ChatGPT isn’t alone in this. If you read no further, take one thing from it: if a chat is private, don’t rate it.

And the chats you’ve already rated? Where they went depends on the app.

Where a thumbs-up sends your chat

The short version: on a personal ChatGPT plan, a rating means the whole conversation may be used for training even after you opt out. On Gemini with Keep Activity off, a rating lifts the same kind of opt-out and brings the last 24 hours of your chats with it.

Claude keeps a rated conversation for up to 5 years, de-linked from your email before it’s used, and Gemini keeps reviewed feedback for up to 3 years, disconnected from your account. Copilot says it doesn’t use ratings to train its foundation models.

ChatGPT

OpenAI’s rule covers personal plans, “For individual plans (Free, Plus, and Pro)”: rate a reply, and “we may use the entire conversation associated with that feedback to improve our models, even if you have opted out of model training” (OpenAI Help, How your data is used to improve model performance). Business, Enterprise and Edu work the other way: there, OpenAI says a rating “does not permit training on the associated conversation.”

Gemini

Switch Keep Activity off and Google says your future chats “won’t be used to train our AI models, unless you choose to send Google feedback.” Clicking Good response or Bad response under an answer is how you give that feedback, and Google adds the conversation to it.

Send feedback with it off, and Google collects your feedback plus “context that can help us better understand your feedback, including the last 24 hours of your chats”. Feedback is “Reviewed by specially trained teams” and kept for up to 3 years, disconnected from your account (Gemini Apps Privacy Hub).

Claude

A rating keeps the whole conversation: “When you provide us feedback via our thumbs up/down button, we will store the entire related conversation, including any content, custom styles or conversation preferences, in our secured back-end for up to 5 years” (Anthropic, Is my data used for model training?). Anthropic de-links feedback from your email before using it, and the uses it lists include training its models “as permitted under applicable laws”.

Copilot

On personal accounts, Microsoft rules ratings out of training for the updated Copilot app: “We don’t use this feedback to train the foundation models used by Copilot” (Microsoft Support, your activity history). It may still use them to improve Copilot.

Across the four

A rating feels like a courtesy to the app. On three of the four, it is also a decision about your conversation.

The rating is the setting nobody sees. The rest sit in menus, and they answer different questions.

Three questions your settings answer separately

Every privacy setting in these apps answers one of three questions. Is the chat in your history? Does the company still have a copy? And what may the company do with it, from training its models to having a person read it?

Most switches answer one. On ChatGPT, the training switch answers the third, and only from then on: OpenAI says that once you opt out, “we won’t use your new conversations to improve our models.” Switching it off deletes nothing.

On ChatGPT and Gemini it also starts on: OpenAI says data sharing is “enabled by default” on personal plans, and Gemini’s Keep Activity is on by default for anyone 18 or over (Google, Manage & delete your activity). Someone chose those defaults before you did, as I wrote in The Setting You Never Changed.

Deleting a chat answers the first question, and only partly the second, because two kinds of copy can outlast it. Memory is one. OpenAI says deleting a chat “does not necessarily delete a separate saved memory created from that chat” (OpenAI Help, Memory in ChatGPT), Claude’s help page says the same of its memory, and Copilot keeps saved memories when you switch its memory off.

For Gemini, Google says to delete the chats that hold it, and where Gemini offers a memory summary, you can ask Gemini to remove things from that too.

Shared links are the other. On ChatGPT and Gemini, anyone who has the link can read the chat; on Claude, anyone with a public link can. Keeping a link out of search engines “does not make a link private”, OpenAI says, and on ChatGPT and Gemini deleting a link leaves the copies people already saved.

Gemini’s Keep Activity is the one setting that reaches all three questions. With it off, new chats stay out of your activity and out of training unless you send feedback, and Gemini’s memory of past chats stops working.

Some uses no setting reaches. Google says that even with Keep Activity off or in a temporary chat, it uses your chats to “help protect Google, our users, and the public, including with help from human reviewers”, and that “Your Gemini settings don’t control processing of your chats to create anonymized data to improve Google services.”

So a chat can be gone from your history and still kept, out of training and still open to a safety review, disconnected from your account and still in front of a reviewer.

Settings tell only part of the story. Who can actually read these conversations, and how long can the companies keep them?

Who can read it, and how long it stays

OpenAI

OpenAI lets a limited number of its staff and trusted service providers access your content for four reasons: investigating abuse or a security incident, support you ask for, legal matters, and improving its models (OpenAI Help, How OpenAI handles data in consumer services). Only the last depends on your training setting. OpenAI says it logs every access.

Anthropic

Anthropic starts from no. By default its employees can’t access your conversations, with two exceptions: you explicitly consent to share them as part of giving feedback, or “Review is needed to enforce our Usage Policy”, when only designated Trust & Safety staff may look, on a need-to-know basis (Anthropic, how it protects Claude users’ data). If you’ve let Claude train on your chats, there is a third: “Access is limited to a small number of personnel involved in model training” (Claude Help, who can view my conversations).

Google

Google has human reviewers, including trained reviewers at its service providers, read some Gemini chats, and disconnects chats from your account before they go to those providers.

Microsoft

Microsoft says that in its updated Copilot app, available since 18 August 2026, chats aren’t used to train its foundation models. The two privacy pages for that app that I read mention neither human review nor how long chats are kept. The older app’s FAQ, which Microsoft now marks as applying “only to the older version”, says some conversations get human review for product improvement and safety, and that conversation activity is kept for 18 months by default (Microsoft Support, Privacy FAQ for Microsoft Copilot).

How long copies last

How long a copy lasts depends on what happened to it, and the principle matters more than any one number: deleting the chat you can see doesn’t necessarily delete every record kept about it.

Three examples show how far apart the clocks run: within 30 days after you delete a chat on ChatGPT or Claude, up to 2 years for a Claude chat its automated safety systems flag as breaking its rules, and up to 3 years for a Gemini chat a reviewer has read, which Google doesn’t delete when you delete your activity.

The 30 days has exceptions. OpenAI keeps a cleared chat it has already de-identified, or must keep for security or legal reasons. Anthropic may keep chats “as required by law, to resolve disputes, or as necessary to combat violations of our Usage Policy” (Anthropic, How long do you store my data?). The private modes keep a copy too: up to 30 days for a ChatGPT temporary chat, 30 days by default for a Claude incognito chat, and 72 hours on Gemini with Keep Activity off.

Ordinary copies can last a while too: Gemini keeps your activity for 18 months unless you change it, and Claude may keep chats you let it train on for up to 5 years in de-identified form.

OpenAI and Google give the same advice. OpenAI: “Please do not enter sensitive information that you would not want reviewed or used.” Google asks you not to enter confidential information “that you wouldn’t want a reviewer to see”. Between them, that is the rule: write as if a person might one day read it.

At work there is a reader who needs no flag and no button.

On a work account

On a work account, your employer runs the account, and on all four apps its administrators or account owners can get at your chats in some form.

OpenAI’s two pages describe ChatGPT Business differently. Its enterprise privacy page says workspace admins “can view, access, export, and delete end user conversations in the workspace” (OpenAI, Enterprise privacy). Its Business help page says each member’s chats aren’t automatically visible to other members, that usage analytics don’t let admins read them, and that “Workspace owners can export a member’s data when the Export data option is available” (OpenAI Help, ChatGPT Business).

So your colleagues don’t see your chats by default, but your employer may be able to view or export them. Don’t count on a work chat being out of your employer’s reach.

On Gemini at work, Google says admins can keep, search and export chats with Vault even when a user’s own action or setting deletes them (Google Workspace Help, Vault).

On Claude Enterprise, the Compliance API pulls chat data (Claude Help), and on Claude’s Team and Enterprise plans incognito chats are included in the exports owners can take; the same plans only let you share chats inside your organisation.

Microsoft’s admins can search Copilot chats with Content search or Purview, and its retention guide warns: “Messages visible in your AI apps are not an accurate reflection of whether they are retained or permanently deleted for compliance requirements” (Microsoft Learn, retention for Copilot and AI apps).

Type nothing into a work AI account that you wouldn’t put in a work email.

Most of this comes down to three things you can do tonight.

The basics

☐ Don’t rate a reply in any chat you want kept private.

☐ Switch training off. ChatGPT: Settings, then Data controls, then “Improve the model for everyone”. Claude: Settings, then Privacy, then “Help Improve our AI models”. Gemini: Activity, then Turn off. Copilot’s updated app says it doesn’t train its foundation models on your chats.

☐ Clear your shared links. ChatGPT: Settings, then Data controls, then Manage next to Shared links, then More actions, then “Delete all shared links”. Claude: Settings, then Privacy, then “Manage” next to Shared chats. Gemini: Settings, then Your public links, then “Delete all links”.

The full check, and what to do about the chats you’ve already rated, follow for paid subscribers.