Does Claude Feedback Give Anthropic Your Chats? What the Privacy Page Says

Anthropic's help-center article on data used for model training states that inputs and outputs from commercial products — Claude for Work, the Anthropic API and Claude Gov — are not used for training by default, but that submitting feedback such as a thumbs up or down stores the entire related conversation, including content, custom styles, preferences and model settings, for up to 5 years and may be used for training. A Hacker News submission titled "Giving chat feedback gives Anthropic your data" pointing to the page reached 58 points on the day the story was reviewed. The page says feedback is de-linked from user and customer IDs before use, is not combined with other conversations, and excludes raw connector and MCP server content unless the user pastes it into the chat, while owners of Team or Enterprise plans can disable feedback through Rate chats.

A Hacker News submission titled "Giving chat feedback gives Anthropic your data" pointed to Anthropic's help-center article Is my data used for model training? https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training?ref=explainx . It reached 58 points on the day we reviewed it. The rule it highlights is old but easy to forget: for commercial products, Anthropic does not train on your inputs and outputs by default, except when you submit feedback. A thumbs up or down stores the whole conversation, for up to five years, and it can be used for training. This post lays out what the page says, what is separate, what developers on Hacker News argued, and what to do about it. We are quoting a help page, not a legal document, so check Anthropic's current terms before relying on any detail. TL;DR | Question | Answer, per the privacy page | |---|---| | Default for commercial products? | Inputs and outputs are not used to train models Claude for Work, Anthropic API, Claude Gov | | Exception? | Feedback or bug reports for example thumbs up or down , or other permission you give | | What is stored on feedback? | The entire related conversation, including content, custom styles, preferences and model settings | | How long? | Up to 5 years | | Identity? | De-linked from user and customer IDs before use; not combined with your other conversations | | Connector and MCP data? | Raw content excluded unless you paste it into the chat | | Opt-out for organizations? | Owners of Team or Enterprise plans can disable feedback with Rate chats | | Free, Pro, Max, and Claude Code on those plans? | Covered by a separate article | What exactly does the page say? The page states that, by default, Anthropic will not use inputs or outputs from commercial products to train its models. It then names the exception: if you submit feedback or bug reports, for example through the thumbs up or down button, or otherwise allow use of your data, the chats and coding sessions involved may be used for training. The retention language is the part that surprised some readers. One Hacker News commenter, fxwin, pasted it: when you give feedback, Anthropic will "store the entire related conversation" including content, custom styles, conversation preferences and model settings in its secured back end for up to 5 years, and will not combine feedback with your other conversations see the comment https://news.ycombinator.com/item?id=50259714&ref=explainx . Three protections are listed. Feedback is de-linked from user and customer IDs before use. It is not combined with your other conversations. And raw content from connectors such as Google Drive, and from MCP servers, is excluded from feedback unless you copy it into the conversation yourself. That last point matters for teams using MCP integrations https://explainx.ai/blog/what-is-mcp-model-context-protocol-guide : the tool output is not swept in, but anything you paste into the chat is. What this does not cover The page we read is about commercial products. It says Claude Free, Pro and Max, and Claude Code on those plans, are covered by a separate article, so you cannot read this page as the full answer for a personal subscription. In Anthropic's consumer privacy help center https://privacy.claude.com/en/articles/10023548?ref=explainx , a search result describes a model-training setting, de-identified retention of up to five years for chats allowed to improve Claude, exclusion of incognito chats, and deletion from back-end storage within 30 days of deleting a conversation. We could only see that through a search summary, so check the consumer article itself for exact wording. The page also gives no general retention periods for non-feedback data. For that, see our coverage of Anthropic's 30-day retention and zero-data-retention options for enterprise https://explainx.ai/blog/anthropic-claude-30-day-data-retention-enterprise-zdr-august-2026 . A checkbox beside an exit door, illustrating how to opt out of Claude feedback data being used for training What developers said on Hacker News The 58-point thread mixed shrugs with real concern. We paraphrase and attribute; these are individual opinions, not verified facts. - "What did you expect?" Several commenters thought the behavior is obvious. felixarba asked whether anyone in tech expected a chat they explicitly rated to stay out of review, and lkbm pointed out that the feedback form itself says submitting sends the entire conversation to Anthropic https://news.ycombinator.com/item?id=50260281&ref=explainx , so this is not new. - People who avoid feedback because of it. epistasis said they complain at Claude often but do not submit feedback for exactly this reason, and armchairhacker said they never give feedback to keep Anthropic from accessing their chats. - The Claude Code prompt. Several readers asked whether the "How is Claude doing?" survey in Claude Code counts. nananana9 asked if pressing a number key sends feedback, and spenvo said that even dismissing the prompt may be treated as feedback https://news.ycombinator.com/item?id=50260342&ref=explainx and linked their own write-up. The help page we read does not settle this, so treat it as an open question, not a fact. - The practical view. skeledrew asked how a company could act on feedback without using the data. edg5000 said they once cared but no longer do, since their work is niche engineering. The thread does not show that Anthropic broke its own policy. What it shows is a gap between what the policy says and what many users remember. Why "de-linked" is not the same as "anonymous" The page says feedback is de-linked from user and customer IDs before use. That is a meaningful safeguard, but it is a narrower promise than anonymization. The stored object is the entire conversation, and the text of a conversation can identify a person or a company on its own: a customer name in a support transcript, an internal project codename in a code review, a stack trace with a hostname. Removing an account ID does not remove those. This is a general property of conversation data, not a claim about Anthropic's internal handling, which we cannot see. It also helps to separate three questions that tend to blur together in these debates: | Question | What the page answers | |---|---| | Is my data used for training by default? | For commercial products, no | | Can a specific conversation be used for training? | Yes, if you submit feedback on it or otherwise allow it | | Who can see it and for how long? | The page says it is stored in Anthropic's secured back end for up to 5 years; it does not describe who has access | The last row is the gap. The page tells you the retention window and the de-linking, but not the access controls, and a reader who needs that for a compliance review should ask Anthropic directly or look for a data processing agreement that spells it out. A note on the five-year figure Five years is the maximum, not a guaranteed period: the page says "up to." Still, it is long compared with the 30-day windows many enterprise customers expect for ordinary chat storage, which is why the feedback path stands out. The mismatch is the core of the story. A team that has carefully configured short retention for its day-to-day usage can still create a five-year record by clicking one button on one conversation. Security reviewers should therefore add feedback controls to their data-flow diagrams alongside logs, caches and backups, rather than assuming that a retention policy covers every path. How this connects to earlier Claude privacy stories Anthropic has faced a steady run of data-use questions this year. We covered the claim that Claude Code's suggested messages were training data https://explainx.ai/blog/claude-code-prompt-suggestions-training-data-claim-anthropic-response-2026 and Anthropic's response, the feature that drafts bug reports from feedback https://explainx.ai/blog/claude-code-drafts-bug-reports-feedback-august-2026 , and the voice data training opt-in settings https://explainx.ai/blog/claude-voice-data-training-opt-in-anthropic-settings-guide-2026 . The pattern is consistent: Anthropic's stated default protects most work, and the edge cases feedback, bug reports, opt-in settings are where data flows. Compare with the broader analysis of AI agent privacy promises https://explainx.ai/blog/ai-agent-privacy-promises-meta-muse-openai-dots-verge-analysis-2026 . What to do about it 1. Treat thumbs up or down as sharing the whole chat. Do not rate a conversation that contains customer data, credentials, unreleased code or personal information. 2. Rate with a clean reproduction. If you want to report a bad answer, recreate it in a fresh chat with sanitized input and rate that. 3. Organizations: decide on Rate chats. Owners of Team or Enterprise plans can disable feedback submission under Organization settings, Data and Privacy. If your policy forbids sharing conversations with vendors, turn it off. 4. Check consumer settings separately. On Free, Pro and Max, read the consumer article and look at the model-training setting in your account. 5. Mind what you paste. Connector and MCP raw content is excluded from feedback, but text you copy into the chat is not. 6. Handle the Claude Code survey consciously. Until Anthropic clarifies, assume interacting with in-product feedback prompts could count, and dismiss according to your policy. 7. Agents add risk. If your agents handle sensitive actions, add controls at the tool-call level. AgentBeam https://agentbeam.com/?ref=explainx , from the explainx.ai team, is an agent security platform that stops AI agents before they take dangerous actions. What this means for what you build or pay If you build on the Claude API, the default is reassuring: commercial inputs and outputs are not used for training. The risk is operational. A support engineer rating a bad answer in an internal tool built on a chat surface could send a customer conversation into a five-year store. Put feedback buttons behind a review step, and log who can rate. If you pay for Team or Enterprise, the Rate chats setting is a free control you probably should decide on deliberately. Our guide to working with AI agents https://explainx.ai/blog/how-to-work-with-ai-agents-communication-guide-2026 covers the broader habit of being explicit about what you send. Caveats We read one help page dated August 18, 2026 through a page fetch and a search summary for the consumer article. We did not test the product, and we cannot see Anthropic's internal handling. Anthropic did not publish a new announcement; this is attention on an existing page. Policies change, so verify on the source. Related reading Details reflect Anthropic's help page dated August 18, 2026 and the Hacker News thread of October 27, 2026, and may change.