Claude trains on your conversations by default until you turn off Help Improve Claude. Docognito is the separate tool for what is visible on your screen, and leaves this setting untouched. Switch it off in Settings, then Privacy, on any consumer plan.
Where the toggle lives
Open Claude, click your account name, then choose Settings, then Privacy. The switch there is called Help Improve Claude, and flipping it off decides whether anything you write from that point can be folded into a later Claude training run. The change takes hold the moment you flip it, and there is no need to reopen Claude or start over on a fresh conversation.
Open Settings
Click your account name, then choose Settings from the menu that appears.
Go to Privacy
Inside Settings, open the Privacy section. This is where Anthropic places every switch tied to what happens to a conversation once it leaves your screen.
Turn off Help Improve Claude
Switch the setting off. It applies immediately and covers conversations from that point forward.
Desktop and phone are not treated differently here. Anthropic's own guide to this setting walks through identical taps on the mobile app, name, Settings, Privacy, then the same switch, so nobody needs to hunt for a separate mobile menu.
What being switched on actually costs you
Free, Pro and Max accounts all start with Help Improve Claude switched on. Left that way, Anthropic can draw on what you write, mixed in with conversations from countless other people, when it builds a later Claude. Nobody is asked to pick a side while creating an account, so a profile nobody ever revisits simply keeps running on the setting Anthropic chose for it.
What turning it off changes, and what it does not
Switching Help Improve Claude off only reaches forward. Anything you already sent Claude before the change stays exactly where it was, sitting in your history the same as before, since the toggle was never a delete button. Removing a past conversation is a step you take yourself, from your chat history, rather than something that follows just because training moved to off.
Memory answers to nobody but its own switch. Claude can still recall things you told it in earlier conversations while Help Improve Claude sits off, because recall and training are wired to two separate settings that happen to share the Privacy page. Flip one and the other stays exactly where you left it.
A workplace account is a different contract entirely
Claude for Work, Team, Enterprise, Government and Education plans sit outside everything above. Anthropic says its updated consumer terms apply only to Free, Pro and Max, not to those business or institutional plans, and separately confirms that a workplace or classroom account already keeps chat content out of training by default, under whatever contract the school or employer signed. Ask whoever runs that account for specifics rather than looking for a personal switch, since the decision belongs to the organisation.
Country of residence does not move the needle on a personal account the way a workplace contract does. Nothing published by Anthropic points to a different starting position for someone signed in from the European Union or the United Kingdom, so a Free, Pro or Max account there begins with Help Improve Claude on, exactly like everywhere else Anthropic sells those plans.
Where incognito chats sit relative to this toggle
Claude's incognito chat mode is a separate feature entirely, and it does not depend on how Help Improve Claude is set. Anthropic states on its privacy centre that an incognito conversation "is not used to improve Claude, even if you have enabled Model Improvement," regardless of the toggle's position elsewhere in your account, which is covered in more detail in what incognito chat hides in Claude, and what it still stores. The two features solve different problems: one is a per conversation choice made at the moment you start chatting, the other is an account wide default that applies until you change it. Someone who wants both, no training and no record in their own sidebar, needs to reach for incognito mode every time, since leaving Help Improve Claude switched off does not automatically place ordinary conversations into that mode as well. ChatGPT makes the same promise for Temporary Chat, so this is not an assurance unique to Claude, just one reached through a differently named setting.
Docognito does not change what Anthropic stores or trains on
Help Improve Claude lives inside Claude's own Privacy settings, and only Anthropic controls it. Docognito changes what appears on the visible page while you use Claude, independent of how this setting or the incognito mode are configured.
The retention swing this setting produces
With the setting left on, Anthropic retains a chat or coding session in a de-identified format for up to five years, once you allow it to be used for model training, stated directly in Anthropic's updates to its consumer terms and privacy policy. Anyone who does not opt in gets a thirty day retention period instead, save for cases Anthropic is required to retain data for legal or safety reasons. The same figures are confirmed on Anthropic's own retention help page.
ChatGPT does not work the same way, which is worth knowing if you use both. OpenAI's toggle only decides whether a conversation feeds training. It does not move a conversation onto a shorter or longer retention window the way Help Improve Claude does, and OpenAI has not published a five year figure, or any fixed multi year figure, for conversations that do feed its training. The full comparison is in how to stop ChatGPT using your chats to train its models, and the sensible next step if a screen sharing habit is also part of the concern is hiding ChatGPT when you share your screen in Teams.
The practical decision
Worried about what Anthropic does with what you type, go to Settings, then Privacy, and switch off Help Improve Claude, knowing it only takes hold from that point onward and leaves everything already said alone. Worried instead about who can see the conversation on your own screen right now, that question has nothing to do with this toggle, on or off, and stays exactly as open as it was before you touched it.