Why this matters
Talking to an AI assistant feels private, like a search you typed and forgot. It is not. By default, the free versions of the big assistants learn from what you type, and a sample of it is read by people.
Training on your chats means your conversations are fed back in to improve the model. Your words, and whatever you have put in them, become part of what the next version is built from. A throwaway question about your health, a draft email with a client’s name, a paragraph of your own writing: all of it is material unless you say otherwise.
This was not always the default, and it is moving one way. When Anthropic, the maker of Claude, updated its consumer terms in August 2025, chats from its free and paid personal plans began training the model unless the user opted out, with that data kept for up to five years. ChatGPT and Gemini sit in the same place: on by default, off by a setting most people never open.
None of this asks you to stop using the tools. It asks two small things of you. Turn the training setting off, and keep what is private out of the box. The rest of this guide is those two things.
Your chats train the model and get read by people unless you say otherwise. Three settings, one per assistant, turn that off.
What they keep
Three things happen to a message you send, and only the first is needed to answer you.
The first is that it is stored. A copy goes into your history, and usually onto the company’s servers for some period beyond that. This is the part you can see, and the part you can often delete.
The second is that it may be reviewed. A sample of conversations is read by trained reviewers to grade whether the answers were good. Google says a subset of Gemini chats are read by human reviewers to improve its generative AI models, and keeps the reviewed ones for up to three years on a path you cannot clear.
The third is that it may be learned. Your words become training data, woven into the next model. This is the one that does not undo. A stored chat you can delete and a reviewer moves on, but a model that has learned from your words keeps them, and there is no way to reach in and take them out again.
Where the off switch is
Each assistant hides the control in a different place and gives it a different name. Here is the one setting that matters in each.
Turn off Improve the model for everyone
Open Settings, then Data Controls, and switch off Improve the model for everyone. It is on by default on personal Free and Plus accounts, and turning it off applies to new conversations across your account. For a single private question, open a Temporary Chat from the top of the chat screen, which is saved nowhere and trains nothing.
Turn off Keep Activity
Go to myactivity.google.com/product/gemini and switch off Keep Activity, the setting once called Gemini Apps Activity. New chats then stop being added to your activity and stop being sampled for human review. You can delete past activity on the same page.
Turn off the improve-Claude setting
Open Settings, then Privacy, and turn off the setting that lets Anthropic use your chats to improve Claude. Until 2025 Claude did not train on consumer conversations. Since the August 2025 change it does unless you opt out, so this is worth checking even if you set it once.
Paid business and developer products, such as ChatGPT Enterprise, Gemini for Workspace, Claude for Work and the APIs, are generally not used to train models by default. The training in this guide is the one switched on across the free consumer apps.
What it does not do
Turning the setting off is worth doing, and it is not a full eraser. It changes what happens next, not what already happened.
Anything already used in a training run stays there. Opting out is forward-looking, so it protects the chats you have from here, not the ones the model has already learned from. A conversation already read by a reviewer can sit on a separate retention path you cannot clear, up to three years in Google’s case.
The service also keeps recent chats for a short while to run at all. Even with activity switched off, Gemini still uses a conversation to produce the reply and to guard against abuse before it ages out. The setting lowers your future exposure. It does not reach into the past, which is the reason the next section matters as much as this one.
Sensible habits
The setting handles the model. Your habits handle everything the setting cannot.
Keep secrets out of the box. Do not paste passwords, card numbers, security keys or other people’s personal details into a chat. If an assistant or an AI agent needs to act on a secret, the pattern in Lock and Key hands the value to the tool without ever showing it on screen.
Use a temporary chat for a sensitive one-off, and write as though a stranger might read the line, because one might. For medical, legal or money questions, ask them in an account you have set to no training, or ask them without naming yourself or anyone else.
Once training is off and a temporary chat is a click away, the careful path is also the easy one. You stop having to decide each time, and the private questions go somewhere that keeps nothing.
Quick reference
| Assistant | Setting | Where |
|---|---|---|
| ChatGPT | Improve the model for everyone, off | Settings, Data Controls |
| ChatGPT, one-off | Temporary Chat | top of the chat screen |
| Gemini | Keep Activity, off | myactivity.google.com/product/gemini |
| Claude | the improve-Claude training setting, off | Settings, Privacy |
| Any | keep secrets and names out of the chat | a habit, not a setting |
Common questions
The questions people ask once they learn their chats were never quite private.
Does turning training off delete what they already have?
No. Opting out stops your future chats from being used. Anything already included in a past training run cannot be pulled back out, because a model that has learned from your words does not unlearn them. Your stored history is a separate thing and can usually be deleted, but that is not the same as removing your words from the model.
If I turn it off, can I still use the assistant normally?
Yes. The setting only governs whether your chats train the model and get reviewed by people. The assistant answers exactly as it did before, and you lose none of its ability.
Does Claude train on my chats now?
Since a terms change in August 2025, Claude uses Free, Pro and Max conversations to improve the model unless you opt out, and keeps that data for up to five years if you allow it. Before that change it did not train on consumer chats. Open Settings, then Privacy, and check the training setting is off if you would rather it did not.
Are my work or business accounts affected?
Generally no. ChatGPT Team and Enterprise, Gemini for Workspace, Claude for Work and the paid developer APIs are not used to train models by default. The training that this guide turns off is on the free consumer apps, which is where most people type.
Can a real person read my chats?
A sample, yes. Each company has reviewers read a portion of conversations to grade the answers. Google keeps human-reviewed Gemini chats for up to three years and does not let you delete those. The safe assumption is that anything you type could be read by a stranger.
What is Temporary Chat or incognito mode?
A conversation that is not saved to your history and not used to train the model. In ChatGPT it is called Temporary Chat, reached from the top of the chat screen. Use it for anything you would rather leave no trace of.
Does deleting a chat remove it from training?
Deleting a chat clears it from your history, but if that chat was already part of a training run or read by a reviewer, deletion does not reach those copies. Turning training off going forward is the part that changes what happens next.
Should I stop using these tools?
There is no need. Turn the training setting off, keep secrets and personal details out of the box, and use a temporary chat for sensitive questions. You keep the tool and give up almost nothing.