Examples

1. 1.

   Anthropic's Claude is trained using Constitutional AI - the model first generates a response, then critiques it against 16 constitutional principles (e.g., 'Choose the less harmful response'), then revises it iteratively.

2. 2.

   Constitutional AI reduced Anthropic's reliance on human-labelled harmful-content data by 90% - the model generates its own critique-revision pairs, scaling safety training without proportional human annotation cost.

3. 3.

   Researchers at DeepMind extended Constitutional AI principles to code generation - using a set of coding principles (correctness, security, efficiency) as a critic to iteratively improve generated code quality.