Anthropic Stores Commercial Chat Feedback for Five Years to Train Future Models
Submitting thumbs up or down feedback on Claude for Work or Claude Code sends your entire session to Anthropic for model training. The data is retained for up to five years, though connectors and remote Model Context Protocol servers remain excluded. Workspace administrators can disable user ratings to protect proprietary code.
Why it matters
Audit your team settings immediately to disable chat ratings before sensitive repository code or debugging context is exported into Anthropic's training sets.
TL;DR
- 01Clicking thumbs up or down retains your full conversation for up to five years for model training.
- 02Local and remote Model Context Protocol server data remains protected unless explicitly pasted into the chat prompt.
- 03Team and Enterprise administrators can turn off the Rate chats toggle to block feedback submissions across the company.
Key facts
- Retention Period
- Up to 5 years
- Default Training Policy
- Excluded for commercial tiers
- Feedback Exception
- Enabled on thumbs up/down click
- MCP Server Data
- Excluded unless pasted into conversation
Commercial Opt-Out Defaults and the Feedback Exception
Anthropic reiterated that under standard commercial terms for Claude for Work, Claude Gov, and the Anthropic API, customer inputs and outputs are never used to train foundation models. However, an explicit user action changes this posture: clicking the thumbs up or thumbs down feedback buttons in the web interface or rating chat sessions flags the conversation for ingestion. Once submitted, Anthropic retains the entire related conversation in secured backend storage for up to five years.
What Data Gets Transferred
The feedback payload is comprehensive. It encompasses the complete dialogue, custom formatting styles, conversation preferences, and specific model settings. Raw connector content from integrated tools like Google Drive and local or remote Model Context Protocol (MCP) servers is excluded by default, but any code or sensitive payload copied directly into the prompt buffer is captured. Anthropic de-links customer IDs from the records, but the content is analyzed for research, user behavior studies, and model training.
Disabling Ratings Across Workspaces
Engineering teams handling proprietary codebases can enforce a strict quarantine. Organization Primary Owners and Owners on Team and Enterprise tiers have access to an administrative toggle. Navigating to Organization settings > Data and Privacy exposes the Rate chats switch, which completely disables feedback submission buttons across the entire organization, preventing inadvertent IP leaks.
✓ When to use
- Use explicit thumbs up or down ratings only when working on sanitized, public, or synthetic test code.
- Keep feedback enabled for internal evaluation accounts dedicated exclusively to testing foundation model quality.
✕ When NOT to use
- Do not leave chat rating enabled on enterprise plans where proprietary source code is pasted into prompts.
- Do not submit bug reports via standard thumbs down buttons if an active session contains sensitive API keys.
What to do today
- Navigate to Organization settings > Data and Privacy in the Claude web console.
- Disable the Rate chats toggle across all Team and Enterprise workspaces.
- Remind developers using Claude Code not to paste raw credentials or proprietary keys into prompts.
What the community says
“Lol people are giving feedback?!?! Talk about Actively training your replacement Just go and poison data on Reddit.”
“That's very deceiving and should warrant a class action law suit in my opinion”