Anthropic's Claude 网页端代码包含一个用于撤回反馈记录的控件
Anthropic 的确认警告称，在该对话退出改进计划之前，同一对话上的其他反馈也必须被撤回。
By Ryan Merket · Published
RUNTIMEWIRE INVESTIGATION — Original analysis
Original reporting by RuntimeWire, based on reverse engineering, testing.
为何重要
反馈可能包含整个对话，并可能被用来改进 Claude。客户端代码显示了针对每条记录的撤回请求，但并未验证账户可用性以及后端对被撤回数据的处理。
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withdrawal of an individual feedback record, and warns that other feedback on the same conversation must also be withdrawn for that conversation to leave the improvement program.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 在其公开提供的 Claude web 客户端中有一个内置工作流程，用于查看过去的反馈并请求撤回单条记录。确认提示警告该对话在与之关联的其他反馈仍然有效期间可能会继续留在改进计划中。代码证明客户端具有此工作流程；但并不能证明哪些账户可以使用它，或在发出请求后 Anthropic 的服务器会如何处理。
Anthropic 描述了其围绕构建可靠、可解释和可引导系统的 AI 研究。该段代码涉及一种更直接的控制：用户在发送 Anthropic 可能用于改进 Claude 的点赞或点踩反馈后会发生什么。
A request attached to one record
Claude web client code 在隐私设置中放置了一个 “Your feedback” 部分。其 Manage 控件会打开一个历史表格，包含日期、评分、原因、产品界面和相关对话。一个单独的功能开关控制是否显示撤回操作，因此显示历史记录和允许撤回在界面上是两个独立的条件。…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.