Glossary · Anthropic

Constitutional AI

Constitutional AI is Anthropic's method for training models to be helpful, harmless, and honest using a written constitution and AI feedback. Definition explained.

By Kadin Nestler · May 28, 2026 · Updated May 28, 2026
On this page

Why constitutional AI exists

Pre-CAI alignment relied on reinforcement learning from human feedback (RLHF), which requires huge numbers of human ratings — slow, expensive, and inconsistent across raters. CAI replaces most of that human labor with AI feedback grounded in an explicit, auditable constitution. The 2022 paper "Constitutional AI: Harmlessness from AI Feedback" introduced the method; Claude is the production model trained with it.

How the training process works

  • Step 1: train a base model to generate responses.
  • Step 2: have the model critique its own responses against the constitution ("does this response violate principle X?").
  • Step 3: have the model revise the response to better match the constitution.
  • Step 4: train a preference model on the original vs revised pairs.
  • Step 5: use that preference model to fine-tune the base model via reinforcement learning.

What is in the constitution

Anthropic's published constitution draws from the UN Universal Declaration of Human Rights, terms of service from major tech platforms, ethical principles from other AI labs, and Anthropic's own research. Principles include "choose the response that is most supportive of life, liberty, and personal security" and "avoid being preachy, obnoxious, or condescending." The constitution is public, which is itself part of the transparency value.

Why this matters for buyers

CAI is one reason Claude tends to be more careful and less prone to comply with harmful requests than some competitor models. For regulated industries (healthcare, legal, financial), that conservatism is usually a feature, not a bug. It also means you can read Anthropic's published constitution and predict, roughly, how the model will behave on edge cases — a level of behavioral transparency unusual in the frontier model market.

What it means for your business

For SMBs in regulated industries (healthcare, legal, financial), a model trained with a transparent constitution is easier to defend in a compliance review than one whose alignment process is a trade secret.

  • AI Safety — AI safety is the field focused on making AI systems behave as intended without harmful side effects. Definition, practical risks, and what SMBs should know.
  • AI Alignment — AI alignment is the problem of making AI systems pursue goals that match human values. Definition, methods, and why it matters for production systems.
  • AI Guardrails — AI guardrails are runtime rules and filters that constrain LLM behavior. Definition, types, and how SMBs should use them in production.
  • Claude Opus — Claude Opus is Anthropic's most capable model, tuned for deep reasoning, long context, and agentic coding. Definition, pricing, and when to use it.
  • AI Ethics — AI ethics is the field examining what AI systems should and should not do, and who decides. Definition, principles, and practical SMB implications.
← Back to the full glossary