Anthropic Safety & Constitutional AI

Worth knowing as background — the cert focuses on building, not safety theory, but a question or two may reference these concepts.

Constitutional AI

A training technique where models critique and revise their own outputs against a written constitution — a set of principles drawn from sources like the UN Declaration of Human Rights, terms of service of other AI labs, and Anthropic's own research.

Pipeline (simplified):

1. Generate a response.
2. Have the model critique that response against the constitution.
3. Revise the response.
4. Use these self-critique pairs as RL training data (RLAIF — RL from AI Feedback).

This produces models that refuse harm without needing massive human-rater feedback loops.

Helpful, Harmless, Honest

The "HHH" framing:

• Helpful — actually solve the user's problem.
• Harmless — refuse harm, but don't over-refuse.
• Honest — admit uncertainty, don't fabricate.

Responsible scaling policy

Anthropic publishes a Responsible Scaling Policy (RSP) committing to evaluate models for dangerous capabilities (bio, cyber, autonomy) before deployment, with mitigations gated by capability level.

See also

• claude-overview