// GLOSSARY -- AGENTIC AI

What is Constitutional AI?

1 min read Updated

Constitutional AI (CAI) is Anthropic's alignment methodology where AI behavior is guided by a written set of principles (a 'constitution') that the model uses to self-evaluate and improve its responses during training.

WHY IT MATTERS

Constitutional AI addresses a fundamental challenge: how do you align AI behavior at scale without labeling millions of examples? CAI's approach: give the model principles (a constitution) and have it critique and revise its own outputs according to those principles.

The process: the model generates responses, then evaluates them against the constitution ('Is this response helpful? Does it avoid harm?'), revises to better satisfy the principles, and this self-revised data is used for RLHF training.

CAI is significant because it reduces dependence on human feedback for every edge case, scales more efficiently, and makes the alignment criteria explicit and auditable — you can read the constitution.

FREQUENTLY ASKED QUESTIONS

What's in the constitution?
Principles about helpfulness, harmlessness, and honesty. Examples: 'Choose the response that is least likely to cause harm,' 'Choose the response that is most helpful while being honest about uncertainty.'
Is CAI better than RLHF?
CAI uses RLHF but with AI-generated feedback based on principles, rather than purely human feedback. It's more scalable and more transparent about alignment criteria.
Can CAI prevent all harmful outputs?
No. Like all alignment techniques, CAI improves behavior probabilistically. Edge cases, novel attacks, and distribution shift can still produce undesired outputs.

FURTHER READING

Enforce policies on every tool call

Intercept is the open-source MCP proxy that enforces YAML policies on AI agent tool calls. No code changes needed.

npx -y @policylayer/intercept
github.com/policylayer/intercept →
// GET IN TOUCH

Have a question or want to learn more? Send us a message.

Message sent.

We'll get back to you soon.