← AI Terminology
Constitutional AI
Constitutional AI (CAI) is an alignment technique developed by Anthropic where a model is given a set of written principles (a "constitution") and trained to critique and revise its own outputs to comply with those principles — reducing the need for human labelling of harmful content.
It is the core training approach behind Anthropic's Claude models.
It is the core training approach behind Anthropic's Claude models.
Why It Matters in AI
Standard RLHF requires humans to label thousands of harmful responses — expensive, inconsistent, and exposure to disturbing content. Constitutional AI replaces much of that with AI self-critique against written principles, scaling safety feedback without scaling human labelling costs. It also makes safety criteria explicit and inspectable rather than implicit in human preferences.
Key Points
| Aspect | Description |
|---|---|
| RLAIF | Reinforcement Learning from AI Feedback — broader term for using AI labels instead of human ones |
| Advantage | Scalable, consistent, interpretable — human values encoded in text rather than implicit in ratings |
| Limitation | Constitution must be well-designed; model may find ways to comply literally but not in spirit |
| Constitution | A document of principles (e.g. "don't assist with bioweapons", "be helpful, honest, harmless") |
| RL-CAI stage | AI-generated preference labels (based on the constitution) used for RLHF — replaces human labelling |
| SL-CAI stage | Model critiques its own responses and revises them against the constitution — supervised learning |
Simple Analogy
Instead of hiring 10,000 moderators to label harmful AI responses, you write a code of conduct and teach the AI to ask itself "does this response violate any principle on the list?" — then revise until it doesn't. The written principles do the work that human labellers would otherwise do.
Common Usage Examples
- Anthropic's Claude: trained with CAI, resulting in a model that discusses controversial topics without refusing to engage
- CAI paper (Bai et al., 2022): first large-scale demonstration that AI self-critique improves safety without human harm labellers
- RLAIF: Google DeepMind and others adopt AI-feedback RLHF as a scalable alternative to human preference labelling
- Constitution contents: "Choose the response that is least likely to contain harmful, unethical, or deceptive content"
- Multi-principle constitutions: Anthropic's includes helpfulness, honesty, and harm-avoidance as distinct axes
Summary
In short: Constitutional AI trains a model to follow written principles by critiquing its own outputs — replacing expensive human labelling with scalable AI self-improvement guided by explicit rules.