Anthropic released a new Constitution for its Claude AI model, published under a Creative Commons CC0 1.0 license. This document replaces a previous list of principles with a detailed explanation of the AI's intended values, reasoning, and behavior. It's written primarily for Claude itself, serving as a foundational guide during training and for generating synthetic data.

The constitution prioritizes four properties: being broadly safe, ethical, compliant with Anthropic's guidelines, and genuinely helpful. It moves away from rigid rules, instead providing context and reasoning to help the model generalize its judgment. This approach, rooted in Anthropic's Constitutional AI training methods since 2023, aims to cultivate better values during model development.

By publishing the constitution, Anthropic seeks greater transparency, allowing users to understand which behaviors are intended versus unintended. The document acknowledges that training models is difficult and outputs may not always align with its ideals. As AI systems gain more influence, such clarity becomes increasingly important for public trust and oversight.