Technique · alignment

Constitutional AI

Training harmless assistants using a written constitution of principles and an AI-generated critique/revision loop rather than human labels for every case.

Origin: Anthropic, 2022-12Read origin paper →Also known as: CAI, Constitutional training

Products deploying

Avg research → prod

First commercial deploy

Deployment timeline

Claude Opus 4.6
Deployed 2026-02-16 · Velocity 3y
“Anthropic's core alignment approach is Constitutional AI, used to train Claude models.”
high
Claude 3
Deployed 2026-02-18 · Velocity 3y
“Claude models are trained using Constitutional AI, Anthropic's proprietary alignment method.”
high
Claude Sonnet 4.6
Deployed 2026-02-25 · Velocity 3y
“Anthropic's core alignment methodology is Constitutional AI, used to train Claude models.”
high
Claude Haiku 4.5
Deployed 2026-02-25 · Velocity 3y
“Anthropic's core alignment technique is Constitutional AI, used to train Claude models.”
high
Claude 4.5
Deployed 2026-02-26 · Velocity 3y
“Anthropic's models are trained using Constitutional AI principles for alignment.”
high
Claude 3.5 Opus
Deployed 2026-03-18 · Velocity 3y
“Anthropic's core alignment approach is Constitutional AI, used across Claude models.”
high
Claude Opus 4.7
Deployed 2026-04-16 · Velocity 3y
“Anthropic's Claude models are explicitly built using Constitutional AI, the company's core alignment methodology.”
high

Prior art

Reinforcement Learning from Human Feedback (RLHF)

Techniques built on this

RLAIF (Reinforcement Learning from AI Feedback)