Anthropic appears to sneak in a legal loophole in Claude Code to train on "feedback" even with training off as in Claude Code's source theres a legal term that counts every "Allow edit" you press as feedback.
retained text
I'm on a Max plan with "help improve Claude" turned off. Every Claude Code binary I have installed (2.1.283 to 2.1.309) starts with this comment:
`// Claude Code is a Beta product per Anthropic's Commercial Terms of Service.`
`// By using Claude Code, you agree that all code acceptance or rejection decisions you make,`
`// and the associated conversations in context, constitute Feedback under Anthropic's Commercial Terms,`
`// and may be used to improve Anthropic's products, including training models.`
Accepting or rejecting an edit is usually just pressing Enter at a permission prompt. Read literally, that makes a normal session "Feedback" that can be trained on, whatever your privacy setting says.
To verify it yourself:
strings ~/.local/share/claude/versions/$(claude --version | cut -d' ' -f1) | grep -A3 "constitute Feedback"
Does this clause override the consumer training opt-out or not?
I contacted Anthropic support about this on 2 September and have had no answer in over a month. Has anyone had a clear answer from Anthropic on this?document fetched 2026-10-24 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xd5ruf/anthropic_appears_to_sneak_in_a_legal_loophole_in