Constitutional AICAI
Teaching AI to critique and revise its own outputs against a written set of principles.
Constitutional AI (Anthropic, 2022) is a training method where a model improves itself by critiquing and revising its outputs against a short list of written principles — a "constitution."
Two-phase process:
- Supervised phase: The model generates a response, then critiques it ("Does this response respect user autonomy?"), then revises it according to the critique. Revisions become training data.
- RL phase: An AI feedback model (RLAIF) rates responses for constitutional compliance, replacing human raters at scale.
Why it matters: Human feedback (RLHF) is expensive, slow, and inconsistent. Constitutional AI scales supervision by using AI to evaluate AI — Anthropic can iterate principles without re-labelling thousands of examples.
The approach is also more transparent: the principles that guide Claude's behaviour are written down and can be examined, unlike the implicit preferences learned from human ratings.
Claude's constitution includes principles around harmlessness, honesty, and helpfulness — with nuanced guidance on how to handle conflicts between them.
In plain terms
Instead of having a teacher grade every essay, students use a published rubric to self-critique and rewrite — then a teaching assistant spot-checks using the same rubric.
Related concepts
Reinforcement Learning from Human Feedback
Training AI to be helpful and safe by learning from human preference ratings.
AI Safety
The field working to ensure AI systems do what humans actually want — now and as they become more capable.
Large Language Model
AI trained on vast text to understand and generate language.
Fine-tuning
Specialising a general AI model on your specific data and style.