Constitutional AI trains a model to police itself against an explicit, written list of principles, drawn from sources like human rights frameworks and platform guidelines, rather than depending entirely on humans labelling thousands of individual examples of good and bad behaviour. The model generates a response, critiques that response against the constitution, then revises it, and this self-critique data is used to fine-tune the model directly. A second stage uses the model’s own preference judgments, guided by the same constitution, in place of some human feedback.
The appeal is scale and consistency: a written constitution can be inspected, debated, and applied identically across millions of interactions, where human labelling is slower, more expensive, and prone to rater disagreement. It doesn’t remove the underlying limitation shared with reinforcement learning from human feedback: the constitution itself is written by people, applied by a model checking its own work, and neither the principles nor the model’s adherence to them come with a formal guarantee that holds outside the situations tested.
Constitutional AI and RLHF are complementary, not competing, techniques, most frontier models today use both. Neither is a substitute for the open control problem: they change how a model’s behaviour is shaped during training, not whether a more capable future model can be proven to stay within any set of constraints at all.