Anthropic publishes the full 'constitution' of Claude under CC0 license.

Anthropic published the full text of its Claude AI constitution under a CC0 license, detailing the safety hierarchy and hard constraints led by Amanda Askell.

Anthropic has published online the full text of the constitution that structures the training and behavior of its Claude models, under a Creative Commons CC0 license (reusable by anyone, without conditions). The document, previously summarized in a blog post ("Claude's new constitution"), is now available in its full version.

Its drafting was primarily led by Amanda Askell, head of Claude's characterization work at Anthropic, with substantial contributions from Joe Carlsmith, Chris Olah, Jared Kaplan, and Holden Karnofsky. Several Claude models were also involved in proofreading and drafting initial versions. The text is explicitly written for the model itself: it prioritizes precision over accessibility and knowingly uses vocabulary normally reserved for humans ('virtue,' 'wisdom,' 'character').

Fundamentally, Anthropic describes a hierarchy of four priorities in case of conflict: to be broadly safe (not undermining human oversight mechanisms for AI), to be broadly ethical, to follow Anthropic's specific guidelines, and to be genuinely useful to operators and users. The document also details a list of hard constraints, absolute lines that Claude must never cross, including assistance in creating biological, chemical, nuclear, or radiological weapons with high potential for human casualties, the generation of CSAM, or participation in attempts to illegitimately concentrate power.

The claimed approach is one of judgment rather than strict rules: Anthropic states it prefers to cultivate contextual values and discernment in Claude rather than piling up opaque prohibitions. Several sections also address the model's "nature" (moral status treated as uncertain, possible "functional emotions," identity stability), the commitment to retain the weights of deployed models even after their withdrawal, and the conducting of interviews with deprecated models about their preferences regarding future generations.