Anthropic's updated data handling policy for commercial customers — Claude for Work, the API, Claude Gov — states plainly that inputs and outputs are not used for model training by default. This is the baseline expectation for enterprise AI, and Anthropic is meeting it, not exceeding it. The real policy lives in the details. The feedback mechanism is where the boundary gets interesting. When a user clicks the thumbs up or thumbs down button on a response, Anthropic stores the entire conversation — including content, custom styles, conversation preferences, and model settings — in secured back-end systems for up to five years. That data can be used for training, research, and studying user behavior. The scope is broad: one click on a feedback button converts an otherwise protected conversation into training-eligible material. Anthopic does de-link feedback from user and customer IDs before internal use, and states it does not combine feedback with other conversations. These are meaningful guardrails. But the asymmetry is stark: a single interaction with a UI element transforms the data's legal and practical status from protected to available. The connector exclusion matters for enterprise customers. Raw content from Google Drive, remote and local MCP servers, and other connectors is explicitly excluded from feedback data — unless that content was directly copied into the conversation. This is a sensible line, but it also means the boundary depends on user behavior, not system architecture. Administrative controls exist. Primary Owners and Owners of Team or Enterprise plans can disable the feedback button entirely through the 'Rate chats' setting under Organization settings > Data and Privacy. This is the correct default for security-conscious organizations, but it requires active configuration — the default leaves the feedback button enabled. The consumer side operates under different rules entirely, with Claude Free, Pro, Max, and Claude Code on those plans subject to a separate policy. This bifurcation is standard in the industry but worth tracking: the commercial guarantee is a contractual relationship, not a technical architecture. What Anthropic has built is a compliant, clearly documented data policy that meets enterprise expectations. It is not a radical privacy commitment. The 5-year retention window on feedback data, the breadth of permitted uses, and the opt-in-by-click mechanism all represent value flowing toward Anthropic's model improvement pipeline. The question is whether enterprise customers read the details or just the headline.
