To stop Claude using your chats for future model training, open your account menu, select Settings, then Privacy, and switch off Help Improve our AI models. Anthropic documents this control for Claude Free, Pro and Max, including Claude Code use through those consumer accounts. Turning it off changes training permission, but it does not erase a conversation from your history or undo training already underway.
That distinction matters when you use an AI assistant for study notes, work drafts or personal questions. There are separate rules for ordinary conversations, deleted chats, feedback and safety reviews. The settings and policy details below were checked against Anthropic's public Privacy Center on 14 November 2026.
| Action | What Anthropic says it changes |
|---|---|
| Turn model improvement off | Previous and new chats are excluded from future model training runs, subject to the stated exceptions. |
| Delete a conversation | It leaves chat history immediately and is deleted from backend storage within 30 days, subject to retention exceptions. |
| Submit feedback | The related conversation can be retained for up to five years and used for training. |
How to turn off Claude model training
Anthropic's official settings instructions give this sequence for a browser or computer. Open the menu associated with your name, choose Settings, select Privacy, then find Help Improve our AI models. Set that switch to off if you do not want to allow the described training use. The help page also describes the same Settings and Privacy route in the app.
The wording of the switch is useful: it concerns improving Anthropic's models. It is not a command to remove saved conversations. If your aim is to withdraw training permission, this is the setting to review. If your aim is to remove a particular conversation from history, deletion is a separate action with its own retention rules.
Anthropic says you can change this preference at any time. Its instructions do not require a paid upgrade to make the choice, and the stated scope includes Free accounts. This guide follows the documented menu path; it does not assume that a particular reader's switch is currently on or off.
Does opting out remove old chats from training?
Anthropic says switching off model improvement stops both previously stored conversations and new chats or coding sessions being used in future training runs. However, data can remain part of training that has already started, and the change does not remove its contribution from models already trained. A future training restriction and reversal of completed training are different outcomes.
The company's consumer data retention explanation adds another detail. When you allow model improvement, de-identified data may remain in training pipelines for up to five years. That period applies to new or resumed conversations after the setting has been enabled. It should not be read as a universal five year storage rule for every chat under every setting.
Deleting a conversation removes it from your visible history immediately, while backend deletion occurs within 30 days under the standard rule. The same page says a deleted chat will not be used to train future models. The qualification about training already in progress still applies, so neither deletion nor changing the switch should be described as making an already trained model forget a conversation.
Feedback and safety reviews have separate rules
Before using a thumbs up or thumbs down button, read the consumer training and feedback policy. Anthropic says feedback can include the entire related conversation, its content, custom styles and conversation preferences. It may retain that feedback conversation for up to five years and use it for research, service analysis and model training.
Anthropic says it separates feedback from the user's identifier before use and does not combine that feedback with the user's other Claude conversations. Even so, the scope is broader than the single response you rate. A rating on a conversation containing personal material therefore deserves attention as a separate data sharing decision.
Safety reviews are another exception. Flagged conversations may be used to improve safeguards, detect harmful material, enforce policies or support safety research. The retention page specifies up to two years for flagged inputs and outputs and up to seven years for trust and safety classification scores. It also allows retention for legal requirements, disputes and policy enforcement. An off training switch does not amount to a promise that no safety processing takes place.
Keep training, history and memory controls separate
For consumer accounts, start by matching the action to your purpose: review the model improvement switch for training permission, and review individual conversations for deletion. Anthropic's help pages distinguish these consumer products from Claude for Work and API services, so avoid applying a business account policy to a personal account without checking its scope.
Related guides explain Claude memory controls, exporting Claude chat history and Claude incognito chat retention. They address different controls. Checking which one you need first makes it easier to make a deliberate privacy choice without mistaking a training preference for complete deletion.
