ChatGPT trains on your conversations by default until you turn off one setting. Doing so stops future training but does not shorten how long OpenAI keeps the conversation, unlike the equivalent Claude setting. Docognito is a separate, screen level privacy layer and touches neither one. Turn it off in Settings, then Data Controls, then Improve the model for everyone.
Where the toggle actually lives
Open ChatGPT, go to Settings, then Data Controls. The relevant switch is labelled Improve the model for everyone, and it governs whether the content of your conversations can be used to train future versions of the model. Turning it off applies going forward from the moment you change it, and there is no need to close the tab or sign out for it to take hold.
Open Settings
Click your account name or avatar, then choose Settings from the menu that appears.
Go to Data Controls
Inside Settings, find the Data Controls section, which groups the toggles that decide how your conversations are used and kept.
Turn off Improve the model for everyone
Switch the toggle off. The change applies immediately and covers conversations from that point forward.
The setting applies to your account as a whole, not to one device or one copy of the app. Turning it off on the web version also covers the mobile app and any other place you sign into the same ChatGPT account, so there is no need to repeat the change on every device you use.
What "on" means by default
On a free or Plus account, this setting starts turned on unless you change it yourself. In practical terms, that means the content of a conversation can be used, alongside conversations from other users, to train and refine future models, unless you have gone into Data Controls and switched it off. Nothing about signing up for either plan requires you to make that choice during onboarding, so an account left at its defaults keeps the setting on by default.
What turning it off changes, and what it does not
Turning the toggle off stops new conversations from being used to train future models. It does not delete conversations that already happened before you changed the setting, and it does not remove them from your chat history. Deleting a specific conversation, or clearing your history entirely, is a separate action available from the same Data Controls area, and it needs to be done on purpose rather than assumed to follow automatically from this one switch.
The toggle also has nothing to do with memory. ChatGPT's memory feature, which lets it recall details you have told it across separate conversations, is controlled by its own setting nearby in the same menu. Turning off training does not turn off memory, and turning off memory does not turn off training. Treat them as two independent switches that happen to live in the same settings page.
Temporary Chat sits outside this toggle entirely. OpenAI states that a conversation held in Temporary Chat is not used to train its models no matter how this setting is configured for the rest of your account, the same protection Anthropic states for an incognito chat on the Claude side. See what temporary chat hides in ChatGPT, and what it still stores for the rest of what that mode does and does not change.
Why this does not shorten how long OpenAI keeps your data
Turning this setting off changes what OpenAI trains on, not how long a conversation sits on its servers. OpenAI's own Data Controls FAQ frames the toggle purely as a training choice, and its chat and file retention policy ties the thirty day figure it does publish to conversations you have deleted, to Temporary Chat, and to copies kept for abuse review, never to an ordinary conversation that stays in your history with training switched off. A conversation left in your history keeps whatever retention your separate history settings give it, which in practice means it stays until you delete it or close your account, regardless of this toggle.
That is a genuine structural difference from how Claude handles the same choice, not a coincidence of wording. Anthropic ties retention directly to its own version of this setting, five years while training is allowed, thirty days once it is switched off, published in Anthropic's updates to its consumer terms. OpenAI has not published any fixed number of years for conversations that do feed training, so stating one here would be a guess rather than a fact. See how to stop Claude using your conversations to train its models for that side of the same question.
Docognito does not change what OpenAI stores or trains on
This setting lives inside ChatGPT itself, and only OpenAI controls it. Docognito changes what is visible on your screen while you use ChatGPT, before the toggle is set and after it, and has no reach into training or retention at all.
Where a Team, Enterprise or regional account differs
Accounts on ChatGPT Team, Business, Enterprise, Edu, or accessed through the API, are excluded from training on conversation content by default, under the business terms those plans are sold on and confirmed on OpenAI's enterprise privacy page, so there is generally nothing to turn off in the way described above. If your ChatGPT access comes through work or school, check with whoever administers that account, since the applicable terms and any admin level controls sit with the organisation rather than with an individual toggle in your own settings.
Location does not change the starting position either. OpenAI's published policies do not describe a different default for this setting based on where you are signed in from, so an account in the European Union or the United Kingdom starts with training turned on by default, the same as an account anywhere else, rather than a regional exception that switches it off automatically.
The practical decision
If your concern is whether OpenAI trains on what you write, go to Settings, then Data Controls, then turn off Improve the model for everyone, and understand that the change applies from that point forward rather than retroactively. If your concern is who can see the conversation on your screen while you are having it, that is a separate question with its own separate answer, and Docognito is built for that second one specifically.