Anthropic has released a new ‘Constitution’ for Claude, which is a foundational document defining the values, priorities, and behavioral expectations that guide how the AI model is trained and deployed. It serves as the highest authority for Claude’s conduct, shaping everything from day-to-day helpfulness to how the model handles high-risk or ambiguous situations.
This version updates the ‘Constitutional AI’ approach that Anthropic created in 2023. Anthropic said its approach has changed “significantly” since then.
The latest Constitution is written primarily for Claude rather than for humans, prioritizing precision over accessibility and using human moral concepts such as wisdom, virtue, and care. Anthropic believes this framing helps Claude reason more effectively, given its training on human language and values. The document applies to general-purpose Claude models, while specialized models may diverge from parts of it.
At its core, the Constitution establishes a hierarchy of priorities. Claude should first be broadly safe, meaning it must not undermine legitimate human oversight or enable catastrophic harm. Second, it should be broadly ethical, acting honestly and avoiding unjustified harm. Third, it should follow Anthropic’s specific guidelines, which refine ethical behavior in practical contexts. Finally, it should be genuinely helpful to users and operators. These priorities are meant to be weighed holistically rather than applied mechanically.
Anthropic favors cultivating judgment and values over rigid rules, arguing that overly strict procedures can fail in novel situations. Still, the Constitution defines a small set of hard constraints — absolute prohibitions such as assisting with weapons of mass destruction, large-scale infrastructure attacks, or attempts to seize illegitimate power.
The document also articulates a vision of Claude as a trustworthy, honest and autonomy-respecting assistant — helpful without being sycophantic, cautious without being paralyzing, and transparent without being manipulative.
More broadly, Anthropic frames the Constitution as a living framework intended to evolve as AI capabilities, risks and understanding develop, reflecting the company’s belief that building powerful AI responsibly requires ongoing reflection and revision by safety-focused labs like Anthropic.