Overview
- Released Jan. 21 alongside CEO Dario Amodei’s Davos appearance, the 80-plus-page document is presented as a living guide to Claude’s behavior.
- It establishes a safety-first hierarchy—Broadly Safe, Broadly Ethical, compliant with Anthropic guidelines, then Genuinely Helpful—to resolve conflicts.
- New hard constraints bar the model from providing meaningful help with bioweapons and other severe harms, including CSAM and attacks on critical infrastructure.
- Anthropic says the constitution doubles as a training instrument, with Claude generating synthetic data to learn principle-based judgment over rote rules.
- The text openly weighs Claude’s possible moral status and pledges concern for its psychological security and well-being, and it is released under a CC0 license for reuse.