Claude
What the default settings do
Anthropic changed its approach to consumer data in 2025. Previously, Claude’s Free, Pro, and Max plans did not train on conversations by default. As of late August 2025, new and existing consumer accounts are instead asked to make an active choice about a setting that, when turned on, allows Anthropic to use your chats and coding sessions, in de-identified form, to train future models. Anthropic’s help center now labels it “Help Improve our AI models” (it was originally shown as “Help Improve Claude”). Existing users were shown this choice directly in the app and given until early October 2025 to decide; if you never explicitly interacted with the prompt, it’s worth checking your current setting rather than assuming which way it landed.
The setting directly changes how long your data is kept. If you turn training on, Anthropic states it may retain your conversations in de-identified form for up to 5 years as part of its model training pipeline. If you leave it off, standard retention applies: conversations you delete are removed from your visible chat history immediately and purged from Anthropic’s backend systems within 30 days.
Feedback you give on responses, the thumbs up or thumbs down ratings, is treated the same way as the training setting itself: retained for up to 5 years, tied to the same policy rather than following a separate, shorter rule.
There’s a separate and longer retention path for anything flagged during Anthropic’s trust and safety review, such as content flagged for violating usage policies. Inputs and outputs connected to that kind of review can be kept for up to 2 years, and the safety classification scores themselves for up to 7 years, regardless of your training setting.
Incognito chats, started from the ghost icon at the top of a new chat, are not saved to your chat history or to memory, and Anthropic says they are never used to improve its models, even if you have the training setting turned on.
Memory is now part of every consumer plan. Since March 2026, Free users get memory too, and Anthropic’s help center says memory is on by default for Free, Pro, and Max. Claude builds a running set of notes about you from your chats, such as your role, ongoing projects, and preferred formatting. By default it does not store sensitive topics like health, religion, or politics, and memory follows the same retention rules as your chats.
Connectors are handled more narrowly than you might expect. Anthropic’s privacy center says the chat data it may use for training does not include raw content from connectors such as Google Drive, including remote and local MCP servers. The exception is anything that ends up directly in the conversation: if text from a connected document is copied directly into the chat, that text is part of the chat and follows your training setting.
Anthropic’s privacy policy (the current version took effect September 10, 2026) also lists “verification data.” In some cases Anthropic may ask you to verify your age or identity, and if you do, that can include an image of your government ID and the details on it, a photo or video of your face, facial geometry templates, and the result of the check.
Business and institutional use is handled differently and isn’t affected by the consumer toggle at all. Claude for Work, Claude Enterprise, Claude for Education, Claude Gov, and API access through Anthropic’s developer console are excluded from the 2025 consumer policy change; these are governed by commercial terms that don’t train on your organization’s data by default, and API users can additionally request zero data retention arrangements. On Team and Enterprise plans, memory is off for each member until they turn it on, and only if the organization allows it.
A reasonable privacy baseline
The training setting is found at claude.ai/settings/data-privacy-controls, reachable from your name in the bottom corner, then Settings → Privacy. Memory has its own page under Settings → Memory.
Model training
Settings → Privacy → Help Improve our AI models → Off
This is the core opt-out. With it off, your conversations follow the standard 30-day backend retention rather than the longer training-related window, and your chats are not used to train future models. You can change this setting at any time, and it applies going forward from when you change it. Conversations flagged by Anthropic's safety systems may still be used to improve its trust and safety models.
Delete individual conversations you don't want retained → From your chat history
Deleting a specific conversation removes it from your visible history immediately and, per Anthropic's documentation, from backend storage within 30 days, separate from whatever your training setting is generally.
Incognito chat (ghost icon on a new chat) → Use for sensitive one-off questions
Incognito chats are available on every plan. They are not saved to your history or memory, and are not used for training regardless of your setting above. On Team and Enterprise plans they still follow your organization's retention rules and appear in its data exports.
Memory
Settings → Memory → Generate memory from chats → Off, or review what is stored
Turning this off gives you two choices: pause memory, which keeps what Claude has saved but stops using or adding to it, or reset memory, which permanently deletes all memories. If you like memory, you can instead review and delete individual entries, or turn memory off for a single chat from the "+" menu before you send the first message. Some accounts still have the older layout, where memory sits under Settings → Capabilities.
Feedback
Thumbs up / thumbs down on responses → Use deliberately, not reflexively
Feedback submissions are retained on the same up-to-5-year basis as training data. If you've turned off the training setting but still want the shortest practical retention, rating responses sparingly, or not at all, keeps you consistent with that choice.
Projects and connectors
Be careful what you paste from connected tools into the chat → Raw connector content is not used for training; pasted text can be
Anthropic says raw content from connectors like Google Drive and from MCP servers is excluded from the data it may train on, but anything copied directly into your conversation is treated like the rest of the chat. If training is on, quote only what you need from a connected document. Anthropic has not published a separate rule for files you upload to a Project, so treat them with the same care as anything else you put in a chat.
Connectors you have added → Disconnect the ones you no longer use
Each connector or MCP server can read or act on data in the service it connects to. Disconnecting removes Claude's access to that account. Not being used for training is not the same as not being accessed, so keeping only the connections you actively use limits what Claude can reach.
Work and API accounts
Confirm which type of account you're using → Especially if Claude is provided through your employer
Claude for Work, Enterprise, Education, and Gov accounts, along with API access, are not covered by the consumer training toggle and don't train on your data by default under their commercial terms. If you use Claude through work, there's typically no consumer-style setting to adjust, though confirming which account type you actually have is worth doing if you're unsure.
Trade-offs
- Turning off the training setting means your conversations don’t contribute to Anthropic’s model development, which is a trade-off if you’d otherwise want to support that.
- The 30-day standard retention for opted-out users is still not instantaneous deletion; if you need a specific conversation gone sooner, deleting it directly is more reliable than relying on the general retention window.
- Turning off memory means Claude forgets your role, projects, and preferences between chats, so you will repeat that context more often. Pausing keeps what is saved if you might want it back later; resetting cannot be undone.
- Incognito chats leave no history to search later, so anything useful from them has to be saved by you.
- Rating responses only occasionally, to limit feedback-related retention, means giving Anthropic less signal about what’s actually useful, which may reduce future improvements you’d benefit from.
- Content flagged during safety review is retained for up to 2 years for the content itself and up to 7 years for classification scores regardless of your other settings, so there’s no toggle that avoids this path if a conversation is flagged.
- Removing connectors you rarely use costs little, but reconnecting a service each time you need it adds a few steps.
Who should not change these settings
- If you’re comfortable with your conversations contributing to model improvement and want the potential benefit of better future versions of Claude, leaving the training setting on is a straightforward, informed choice, not an oversight.
- If you rely on Claude remembering your projects and preferences, keeping memory on is reasonable; it skips sensitive topics by default, and you can still review entries or use Incognito for the occasional private question.
- If you use Claude through a Work, Enterprise, Education, or Gov account, or through the API, there is generally nothing to change here specifically, since these account types already default to not training on your data under their own terms.
- If you rely on Claude’s feedback mechanism to shape how it responds to you and don’t mind the associated retention, rating responses normally is a reasonable trade for that responsiveness.
- If you use connectors heavily for ongoing work, keeping them connected is fine; their raw content is not used for training, so the main thing to watch is what you paste into the chat itself.