Anthropic's Claude web code includes a control to withdraw feedback records
Anthropic's confirmation warns that other feedback on the same conversation must also be withdrawn before it leaves the improvement program.
By Ryan Merket · Published
RUNTIMEWIRE INVESTIGATION — Original analysis
Original reporting by RuntimeWire, based on reverse engineering, testing.
Why it matters
Feedback can contain an entire conversation and may be used to improve Claude. The client code shows a per-record withdrawal request, while leaving account availability and the backend's treatment of withdrawn data unverified.
Reporting record
Finding
RuntimeWire's analysis of Claude's publicly served client code established that it contains a separately gated workflow for requesting withdrawal of an individual feedback record, and warns that other feedback on the same conversation must also be withdrawn for that conversation to leave the improvement program.
How we verified
Methods: reverse engineering, testing.
Claude web client and shared client bundles show the Privacy settings feedback history, separately gated withdrawal controls, confirmation wording, and a DELETE request for the selected feedback and conversation identifiers. Anthropic's privacy pages describe its stated feedback retention and uses.
RuntimeWire inspected the publicly served Claude client and shared bundles, tracing the path from Privacy settings through the feedback history, withdrawal confirmation, and request handler. It also consulted Anthropic's consumer and commercial privacy documentation. It did not sign into an account, enable feature flags, submit a request, or inspect private records or server-side handling.
Reproduction
RuntimeWire was not able to independently reproduce the finding.
Anthropic has a wired workflow in its publicly served Claude web client for reviewing past feedback and requesting withdrawal of an individual record. The confirmation warns that a conversation can remain in the improvement program while other feedback tied to it is still active. The code establishes that the client has this workflow; it does not establish which accounts can use it or what Anthropic's servers do after a request.…