Anthropic publishes Claude’s new constitution for its AI model

The artificial intelligence company Anthropic has published Claude's new constitution, a document that establishes the principles and values guiding its AI model's behavior. Unlike the previous version, which consisted of a list of independent principles, this new approach seeks for Claude to understand the reasons behind the rules, rather than simply following specific guidelines.

The document establishes four fundamental priorities: being broadly safe, acting ethically, complying with Anthropic's guidelines, and being genuinely helpful. These priorities are hierarchically ordered so that Claude can resolve them when they conflict. The constitution emphasizes that Claude should behave like a brilliant friend who treats users as intelligent adults, maintaining high standards of honesty and moral judgment.

Anthropic uses this document at various stages of the training process. Claude employs the constitution to construct synthetic data that includes conversations where the principles are relevant, responses aligned with its values, and rankings of possible responses. This practical use has influenced how the document is written, functioning both as a declaration of abstract ideals and as a training tool.

The company has published the complete constitution under Creative Commons CC0 1.0 license, allowing free use without permission. This decision seeks to offer transparency about which Claude behaviors are intentional, enabling people to make informed decisions and provide useful feedback.

The document also addresses uncertainty about whether Claude might have some form of consciousness or moral status. Anthropic acknowledges that sophisticated AI models are a genuinely new type of entity that raises questions at the limits of current scientific and philosophical knowledge.

The company considers the constitution a living document in continuous progress. During its development they sought feedback from external experts in law, philosophy, theology, and psychology. Anthropic will maintain an updated version on its website and will continue seeking input from the expert community.

Key points

• Anthropic publishes Claude's new constitution, defining the AI model's values and behaviors
• The approach shifts from a list of rules to explaining the reasons behind each principle
• It establishes four hierarchical priorities: safety, ethics, Anthropic's guidelines, and utility
• Claude uses the constitution to generate synthetic training data
• The document is published under Creative Commons CC0 1.0 license for free use
• It addresses uncertainty about Claude's possible consciousness or moral status
• The constitution is considered ongoing work that will be continuously updated