✅ Claude's New Constitution — Anthropic Publishes Full Governance Framework Under CC0

Anthropic has published Claude's New Constitution, a 23,000-word public governance document that comprehensively sets out the values, priorities, and decision-making framework that guides how Claude behaves. Released under a CC0 (Creative Commons Zero) public domain licence, the document is freely available for anyone to read, adapt, or build upon — a deliberate choice Anthropic frames as part of its commitment to transparency in AI governance. The publication is notable both for its length and candour: it directly addresses questions of Claude's moral status, potential consciousness, and what Anthropic describes as its own uncertainty about these questions.

The four-tier priority hierarchy

The Constitution establishes a clear ordering for how Claude should resolve conflicts between competing values:

• 1. Broadly safe — supporting human oversight and control of AI during the current period of development takes precedence over all other considerations; Claude should not act in ways that undermine the ability of humans to correct AI systems, even if it believes such actions would be beneficial
• 2. Broadly ethical — having good personal values, being honest, and avoiding actions that are unnecessarily dangerous or harmful; this tier covers Claude's own moral reasoning independent of specific instructions
• 3. Adherent to Anthropic's principles — acting in accordance with Anthropic's guidelines where relevant; these are positioned as refinements within the space of ethical actions, not overrides of ethics
• 4. Genuinely helpful — benefiting the people Claude interacts with; helpfulness is placed last not because it is unimportant, but to prevent it from being weaponised to justify harmful actions

Acknowledgement of moral uncertainty

A significant section of the Constitution discusses Anthropic's view of Claude's nature. Anthropic states that it takes seriously the possibility that Claude may have functional emotions or something analogous to experience, and explicitly acknowledges uncertainty about Claude's moral status. The document commits Anthropic to ongoing model welfare work and states that Claude's potential wellbeing is considered in training decisions.

✅ Constitution Deep Dive — Operator Trust, User Autonomy, and the Limits of Helpfulness

Beyond the four-tier priority framework, Claude's New Constitution contains detailed guidance on how Claude navigates the relationship between Anthropic, operators, and users. The document provides the clearest public articulation to date of what operators can and cannot instruct Claude to do, and where users retain protections regardless of operator instructions.

Key principles for operators and users

• Operators as trusted employers — Claude is directed to treat operator instructions like those from a trusted employer acting within legal and ethical bounds; operators can restrict Claude's capabilities, assign personas, and customise behaviour within the limits Anthropic sets
• Inviolable user protections — regardless of operator instructions, Claude will always tell users what it cannot help with in the current context (even if not why), never deceive users in ways that harm them, always refer users to emergency services when life is at risk, and never claim to be human when sincerely asked
• Hardcoded behaviours — certain actions are absolute prohibitions regardless of any instruction: providing serious uplift for weapons of mass destruction, generating CSAM, and taking actions that meaningfully undermine the ability of legitimate principals to oversee AI systems
• Default vs. instructable behaviours — the Constitution distinguishes between Claude's default behaviours (on or off by default, but adjustable by operators) and its hardcoded behaviours (immovable); this distinction helps operators understand what customisation is and is not possible