Anthropic has released a new constitution for its AI model, Claude, outlining the values and behavior expected from it. This document plays a significant role in the training process, aiming to shape Claude's outputs and ensure they align with the company's ethical standards. The constitution includes guidelines on being helpful, safe, and ethical while allowing Claude to navigate complex situations and trade-offs, such as balancing honesty with compassion. It’s primarily designed for Claude’s understanding, making it a practical framework for its behavior rather than a rigid set of rules. The constitution emphasizes four main priorities for Claude: broad safety, ethical behavior, compliance with Anthropic's guidelines, and genuine helpfulness. In cases of conflict among these priorities, Claude is instructed to follow them in that specific order. The document also incorporates a historical perspective on how Claude is meant to evolve as AI technology progresses. It acknowledges the challenges of training AI, recognizing that Claude's behavior may not always meet the ideals laid out in the constitution. Anthropic's previous constitution consisted of standalone principles, but this new version aims for deeper understanding. It explains the rationale behind desired behaviors, allowing Claude to apply broad principles in diverse situations instead of strictly following specific rules. The document is intended to be transparent for users and developers, helping them provide informed feedback. Key sections detail how Claude should measure helpfulness, adhere to Anthropic's guidelines, maintain high ethical standards, and prioritize safety in its operations.