Constitutional AI & Alignment - Interactive Visualization

Constitutional AI (CAI) trains helpful and harmless models using self-critique guided by explicit principles - no human labelers needed for harm avoidance. The model critiques its own responses against a constitution of principles, revises them, then trains on this synthetic preference data via RL. This demo walks through each pipeline stage.

• Constitution viewer - browse the explicit principles used to guide critique (e.g., non-deception, non-manipulation)
• Self-critique loop - see how the model evaluates its own response against each constitutional principle
• Revision step - watch the model rewrite a harmful response into one that satisfies the constitution
• RLAIF pipeline - understand how the revised responses form the preference dataset for RL training without human raters
• Compare CAI-trained vs base model responses to the same adversarial prompt to see the alignment difference

Part of the EngineersOfAI Interactive 3D - free interactive visualizations covering every major concept in machine learning and AI engineering. Hover any element for a plain-English explanation. No code required.