The four best-known chatbots — ChatGPT, Claude, Gemini and Grok — each have a switch that stops them training on your chats. Despite near-identical names, the switches work in surprisingly different ways. Three of the four are on by default, and three of the four have a significant exception the name does not suggest.

Here is where each switch is, and — more usefully — what it leaves behind.

ChatGPT

setting   Improve the model for everyone
where     Profile → Settings → Data Controls
default   on for Free, Plus and Pro

Switch it off and new conversations stop being used for training. Your chat history stays where it is; this is not the same as deleting anything, and the setting applies to the whole account rather than the device you changed it on.

The default depends on your account type. Business, Enterprise, Edu and the API do not train on your inputs by default. The consumer version, which is the one most people use, does.

Claude

setting   the model training choice in Privacy Settings
where     claude.ai/settings/data-privacy-controls
default   you were made to choose

Anthropic asked existing users to make a decision rather than defaulting them in, which is a better shape than most. The toggle does two jobs, though, and the second is easy to miss.

It also sets how long your chats are kept. Leave training on and they are kept for five years. Turn it off and it is thirty days.

Gemini

setting   Keep Activity
where     Gemini Apps Activity, in your Google Account
default   on

The main trap here is that turning Keep Activity off does not stop human review. Google is explicit that chats are still used to respond to you and to protect the service, including with help from human reviewers, whatever the setting says. With the setting off, chats are held for 72 hours.

Chats a human reviewer has seen are kept for up to three years, and deleting your activity does not delete them. The delete button covers your activity history. It does not reach the reviewed copy.

Grok

setting   Data Sharing, under Grok & Third-party Collaborators
where     X settings → Privacy & Safety → Data sharing and personalization
default   on

Grok's toggle covers more than your chats: your public posts, your profile, and your interactions with Grok all feed the same setting. Making your account private is the other lever X names.

X's own help pages name two exceptions to the opt-out.

Rating a reply opts that conversation back in. Press thumbs up or thumbs down on a Grok answer and X and xAI may use that conversation for training, even with the main switch off. X's advice is blunt: if you object, do not submit the feedback.

And the opt-out does not stop a deployed model learning from normal use of X features that Grok powers, such as recommendations. That is more limited than training a new model, but it is still a form of learning that the name on the switch does not imply.

This is also separate from the Grok Personalization toggle in the same screen; turning one off does not affect the other.

What actually follows from this

Check the default rather than assuming it. Three of these four are on until you change them. Nobody will tell you which state yours is in.

Read what the toggle is scoped to, not what it is called. Every one of these is named as though it governs the same thing. One of them also controls retention, one does not touch human review, and one is undone by a thumbs-up.

Opting out is not deletion. None of these settings removes what was already sent. They govern what happens next. If you want the earlier material gone, that is a separate request through each provider's privacy channel, and in Gemini's case the reviewed copies are outside it.

If it is work material, use the work account. The business and enterprise tiers of ChatGPT and Gemini do not train on inputs by default. That is a far more reliable control than remembering a toggle on a personal login, and it is the one change that survives someone else using the same machine.

The switch is not a secrecy guarantee. Turning off training does not mean nobody can read the conversation. Every provider retains some ability to look at content for abuse, safety and legal reasons. Treat anything you would not want a stranger to read as unsuitable for a chatbot, whatever the settings say.

Where this comes from, and what will date

Each setting was read from the provider's own documentation on 29 August 2026: OpenAI's Help Center, Anthropic's Privacy Center and consumer terms update, Google's Gemini Apps Privacy Hub, and X's own "About Grok" page. The Grok wording about feedback and about deployed models is X's, not a summary of it.

This will go out of date, and faster than most things here. These settings have moved more than once in the past two years — names, defaults and locations. Treat the paths above as where to look rather than as a promise of what you will find, and check the date on anything else you read about them.

Four providers is not a survey. Copilot, Meta AI, DeepSeek, Perplexity and the rest each have their own arrangement. The pattern to carry over is the question to ask, not this list.

Consumer tiers only. Enterprise agreements, API terms and regional variations — the EU in particular is treated differently by several of these — are outside what is described here.

Nothing here was tested by experiment. These are documented behaviours, taken from the people who run the services. There is no way from outside to verify what a provider does with a conversation after it arrives, which is the honest limit on every guide of this kind, including this one.