What is Constitutional AI?

How Constitutional AI Works

Constitutional AI has two phases. First, a supervised phase where the model critiques and revises its own outputs based on constitutional principles ("avoid harm," "be honest," etc.). Second, a reinforcement learning phase where the model is trained to prefer responses that better adhere to the constitution, using AI feedback rather than human feedback. The approach combines RLHF with self-supervision, scaling alignment work without proportionally scaling human labeling.

Why Constitutional AI Matters

Constitutional AI is the alignment approach behind Claude. It addresses a key bottleneck in alignment research: human feedback is expensive and inconsistent. By using AI feedback guided by clear principles, the approach scales better and produces more consistent behavior. Anthropic has open-sourced parts of the methodology, and similar approaches are emerging at other labs.

Practical Example

Claude's training uses constitutional AI principles to balance helpfulness, harmlessness, and honesty. When asked questions where these principles tension against each other (e.g., a request for information that could be harmful), Claude's training produces measured responses that explain the tradeoff rather than refusing entirely or complying without thought.

Use Cases

• AI safety alignment
• Instruction following
• Reducing harmful outputs
• Behavior consistency

Salary Impact

Where this skill pays off

Related Terms

Concepts that pair with this one. Each links to a deep explainer.

Related Skills

Frequently Asked Questions