Last read: 24 Oct 2026 · From Anthropic’s own pages
Claude: what happens to your chats
1. Training
Setting: Settings > Privacy > “Help Improve our AI models” toggle, on web and mobile.
Incognito chats are not used for training.
2. Retention and deletion
Deleted chats leave your history at once and Anthropic’s backend storage within 30 days.
With training on, de-identified data may be kept up to 5 years.
Chats flagged for policy violations: inputs and outputs kept up to 2 years, classifier scores up to 7 years. Feedback data: kept 5 years.
3. Human review
Not stated on the pages we read. They refer to automated trust and safety systems.
What still applies with training off
- Chats flagged by safety classifiers can still be used for safety models.
- Training runs that have already started keep your data.
Worth knowing, in plain words
Terms in a privacy policy can be hard to read. These are the points that matter most for Claude, with a generic example of each. The examples are illustrations, not real cases.
Delete is not instant
Deleting a chat usually removes it from your view first. Removal from the company’s systems follows later, within the period stated.
Example, illustrative onlyYou delete a chat on Monday. It is gone from your list at once, but the stated deletion period can run for weeks after that.
What this page says: Deleted chats leave backend storage within 30 days.
Feedback is a separate channel
Rating or reporting an answer can send that conversation, or a slice of it, to the company, even when your other settings are off.
Example, illustrative onlyYou switch training off, then tap thumbs-down on a bad answer to help the company fix it. That one tap may be treated as permission for that conversation.
What this page says: Feedback data is kept for 5 years.
“Not stated” is not “No”
When the pages we read are silent on a point, we say “Not stated”. That means we found no statement, not that the practice does not exist.
Example, illustrative onlyA page says nothing about human review. You cannot conclude that nobody reads chats; you can only conclude the page does not say.
What this page says: Human review is not stated on the pages we read.
Sources
- Anthropic: How do I change my model improvement privacy settings?
- Anthropic: How long do you store my data? · Updated 1 Jul 2026
“Not stated” means the pages we read are silent on the point. It is not a statement that the practice does not exist. Wording is checked against the page on the date above and may have changed since.
Spot an error? Email corrections@aichatprivacy.com with a link to the source.