Anthropic has updated Claude’s constitution, aiming to improve how its language model engages with users. This approach contrasts with previous reliance on human feedback, where workers would compare model outputs and select the better response. This method has drawbacks: it can involve harmful content, lacks scalability, and demands significant time and resources. The new model, based on Constitutional AI, uses a set of predefined principles to guide output evaluations. This allows Claude to avoid toxic or discriminatory responses and to promote helpfulness and honesty without heavy human oversight. Constitutional AI trains the model through two phases. First, it critiques its own outputs based on the constitution. In the second phase, it uses AI-generated feedback to reinforce learning, focusing on harmlessness rather than human input. This method has shown promising results, with Claude demonstrating improved responses to adversarial prompts while maintaining helpfulness. The approach enhances transparency, making it easier to understand the principles guiding the AI, and reduces the need for human review of distressing content. Claude's constitution is still a work in progress, drawing from diverse sources like the UN Declaration of Human Rights and safety best practices from other AI labs. The team emphasizes the need for ongoing feedback and iteration in shaping these principles. They aim to include a wider range of perspectives beyond Western viewpoints, addressing modern issues like data privacy. The principles developed through trial and error focus on ethical responses while avoiding condescension or excessive judgment, reflecting a commitment to creating a more balanced and thoughtful AI system.