Constitutional AI is an alignment approach developed by Anthropic in which a language model is trained to evaluate and revise its own outputs according to a written set of principles called the constitution. During a self-critique phase, the model is prompted to identify ways a given response might violate those principles and then to rewrite it so the output better conforms; this process generates synthetic preference data that is then used in reinforcement learning, reducing dependence on large volumes of human-labeled feedback. The goal is a more transparent, scalable method for instilling consistent values into large models.

Think of it like hiring a lawyer who has memorized a detailed code of conduct and is required to review every letter they write before sending it. Instead of waiting for a supervisor to flag problems, the lawyer self-edits in advance against the known rules.

In practice, the self-critique loop works as follows: the model first generates a draft response, then receives a follow-up prompt asking it to identify any principle the response might breach, and finally produces a revised version. This revised-response dataset is used to train a preference model, which then guides further RLHF-style optimization without requiring human raters to label every harmful output.

Constitutional AI sits at the intersection of RLHF, Fine-tuning, and broader AI safety research. Related concepts include Hallucination (which the constitution can be designed to reduce), Prompt Engineering (since the principles are themselves textual), and Foundation Model alignment strategies.

You will encounter this term most often in discussions about Anthropic's Claude model family, AI safety literature, and engineering blogs that compare alignment techniques across different labs.