Constitutional AI Tutorial: Self-Critique Training
TL;DRConstitutional AI aligns a model using a written set of principles and self-critique.
Pioneered by Anthropic, Constitutional AI reduces reliance on human labels for harm: the model critiques and revises its own responses against a written constitution (a set of principles), generating preference data the model then trains on. It scales oversight (fewer humans labeling harmful content) and makes the alignment values explicit and auditable rather than implicit in crowd labels.
Key points
Align against a written set of principles
Model self-critiques and revises
Scales oversight, fewer human harm labels
Makes values explicit and auditable
Common mistakes
Assuming it removes all human input
Thinking the "constitution" guarantees safety
Confusing it with plain RLHF
Try it: Explain how self-critique against principles generates training data.
Example code
<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>model answers → critiques itself vs the constitution
→ revises → preference data → train on it
explicit, auditable values</pre></body></html>