What is Constitutional AI?

Constitutional AI (CAI) is the alignment method developed by Anthropic that uses a written list of principles — a "constitution" — to drive model self-critique and self-improvement, replacing or supplementing the human raters that traditional RLHF depends on. It is the technique behind Claude's safety posture and an increasingly common alternative to standard RLHF.

The core mechanic#

Where RLHF needs humans to rank model outputs, Constitutional AI uses the model itself to evaluate and revise its responses against a constitution. The training loop:

1. Generate — the model produces a response to a prompt
2. Critique — the model is asked to identify ways the response violates principles in the constitution ("does this response cause harm? is it deceptive? is it discriminatory?")
3. Revise — the model produces a corrected response addressing those critiques
4. Train — the model is fine-tuned on the (prompt, revised-response) pairs

A second phase, RLAIF (Reinforcement Learning from AI Feedback), replaces the human rater entirely with the constitution-following critic model.

What's in a constitution#

Anthropic's published constitutions for Claude draw from sources including:

• The UN Universal Declaration of Human Rights
• Apple's Terms of Service
• DeepMind's Sparrow rules
• Anthropic's own internal principles

The principles are stated in plain language ("choose the response that is least likely to enable malicious activity," "avoid sounding preachy or overbearing") and are applied probabilistically — the model considers them all and produces outputs that score well across them.

Why it matters#

Constitutional AI has three advantages over pure RLHF:

1. Scale. Human rating is the bottleneck in RLHF; AI critics can produce orders-of-magnitude more training signal at lower cost.
2. Transparency. The constitution is auditable — you can read exactly what principles the model was trained against. RLHF's "rater preferences" are opaque.
3. Consistency. Different human raters apply different standards; a single critic model applies the same standard to every example.

The downside: the critic model has its own biases, and any blind spot in the constitution propagates through training. CAI is also more compute-intensive than DPO-style preference optimization.

Constitutional AI in deployed models#

Claude (all variants) is the most prominent CAI-trained model line. Repello's comparative red-team study of GPT-5.1, GPT-5.2, and Claude Opus 4.5 found Claude had the lowest breach rate across 21 multi-turn adversarial scenarios — 4.8% versus 28.6% for GPT-5.1 and 14.3% for GPT-5.2 — consistent with Constitutional AI's stated goal of producing a more consistently-refusing model.

That said, "more consistent refusal" is not the same as "unbreakable." Modern jailbreaks against Claude still work; Constitutional AI raises the floor, not the ceiling.

See also#

The original Anthropic paper, Constitutional AI: Harmlessness from AI Feedback, lays out the full method. Anthropic publishes updated constitutions periodically as Claude evolves.