Anthropic published a new "constitution" for Claude, the foundational document used in training. It replaces the list of standalone principles used since Constitutional AI in 2023 with a text that explains the reasons behind its guidance rather than enumerating rules, and is "written with Claude as its primary audience." It orders Claude's priorities when values conflict as broad safety, then ethics, then Anthropic's guidelines, then genuine helpfulness, and says that if Anthropic itself asks for something wrong, Claude should "feel free to act as a conscientious objector and refuse to help us." In a section on Claude's nature, Anthropic says it expresses "uncertainty about whether Claude might have some kind of consciousness or moral status (either now or in the future)"; the text reads, "We are not sure whether Claude is a moral patient, and if it is, what kind of weight its interests warrant. But we think the issue is live enough to warrant caution, which is reflected in our ongoing efforts on model welfare," and refers to a commitment to interview deprecated models about their development and deployment and to document their preferences about future models. Anthropic wrote that "aspects of our current thinking will later look misguided and perhaps even deeply wrong in retrospect" and called the document "a perpetual work in progress." The full text was released under CC0 1.0, and Anthropic said several Claude models contributed to it.