Constitutional AI
- Training
- Standard term
Constitutional AI is a training approach where a written list of principles guides an AI as it critiques, revises, and compares answers, and that feedback is used to train the model.
People who rate answers follow written guidelines, but the model learns the preference their thousands of ratings add up to, and that sum is written nowhere. A constitution puts the target in plain sentences that anyone can quote and dispute, and rewording a sentence changes what the next round of training aims for.
Where you’ll meet it
A news story reports that an AI company has published the full written constitution its assistant is trained on. A comment under the story says this means the assistant will now always follow those principles.
What to do
Treat the document as the company's stated aim, a fair guide to what to expect and something to quote when an answer falls short. Training leaves gaps nobody foresaw, so the answers themselves, especially where two principles pull apart, show how far the aim was met.
Easy to mix up
Whoever sets up an assistant can edit its system prompt in the settings. Constitutional AI shapes the model during training, so only the people who train it can change those principles.
Check yourself
Two labs train copies of the same starting model with the same written constitution. Each lab uses a different AI to do the comparisons. Will the two trained models behave alike?
Go deeperIn the original method, people still judged helpfulness
Bai and colleagues at Anthropic described the method in a 2022 paper, “Constitutional AI: Harmlessness from AI Feedback.” In its first stage, each critique and revision drew one principle at random from the list. People still rated answers for helpfulness, so AI feedback stood in for them only on harmlessness, an early case of reinforcement learning from AI feedback (RLAIF).
A model's answer rarely traces back to one clause, because training blends the principles into everything else the model learned. Other labs publish similar documents. OpenAI's Model Spec describes how its models are meant to behave, and its deliberative alignment work trains models to reason over a written safety policy before answering.