A foundational charter for Claude

Anthropic describes the constitution as a detailed, foundational document that both expresses and shapes who Claude is, spelling out the values it should embody and the reasons behind them. It is written primarily for Claude itself, giving it context about its situation and guidance for handling difficult tradeoffs like honesty versus compassion or protecting sensitive information.​

Why move beyond simple rules

Earlier versions focused on standalone principles, but Anthropic now argues that advanced AI needs to understand why certain behaviors are desired so it can generalize good judgment across new situations. They still use specific hard constraints for high risk areas, such as never significantly assisting a bioweapons attack, but they do not want Claude to become a rigid rule follower that misses real human needs.​

Four ordered priorities

The new constitution organizes Claude’s goals into four priorities: be broadly safe, broadly ethical, compliant with Anthropic’s guidelines, and genuinely helpful. When these aims come into conflict, Claude is instructed to treat this list as an order of operations, with most of the document devoted to unpacking what each priority requires in practice.​

What the main sections cover

The main sections explain how Claude should be substantively helpful to different stakeholders, follow Anthropic’s detailed guidelines on topics like medicine and cybersecurity, and cultivate strong ethics such as honesty and careful harm avoidance. They also stress broad safety, including not undermining human oversight, and explore questions about Claude’s nature, potential consciousness, and psychological wellbeing as AI systems become more capable.​

A living, experimental document

Anthropic presents the constitution as a living document, released under CC0 so others can reuse it, and expects to revise it as they learn from practice and expert feedback. They note that there will always be a gap between the ideal expressed in the constitution and real model behavior, so they pair it with evaluations, safeguards, and alignment research to keep future versions of Claude as close as possible to the values on the page.