Claude's New Constitution

Anthropic has unveiled a new Constitution that defines the values and behavior of its AI model Claude. This document is a comprehensive guideline containing the vision and context for what Claude should be.

The new constitution goes beyond a simple list of principles, explaining in detail the reasons and intentions behind requiring specific behaviors. This is meant to help the model go beyond mechanically following rules, fostering the generalization ability to apply principles and make correct judgments even in new situations.

The constitution serves as a core tool for model training. Through Constitutional AI technology, it shapes Claude's values, and it is also used for Claude to understand the constitution itself and generate synthetic data aligned with it or rank responses accordingly.

The document also plays a role in increasing transparency. Through the constitution, users can determine whether a specific behavior of Claude is intentional, and based on this, provide more useful feedback.

This revision focuses on equipping Claude with the ability to make its own judgments in complex trade-off situations, such as balancing honesty and compassion with protection of sensitive information.