Constitutional AI in plain English.

Constitutional AI is Anthropic's technique for training a model to follow a written set of principles — its "constitution" — rather than relying purely on human raters to judge every response, the way traditional RLHF does. Instead of needing thousands of human-labeled comparisons, the method has the model critique and revise its own draft answers against the constitution's rules (for example, "choose the response that is more helpful and less harmful"), generating its own training signal at far greater scale than humans reviewing individual outputs could provide. Anthropic introduced the approach in 2022 and has kept refining it through later Claude releases, publishing successive versions of the constitution itself. The appeal is scalability and transparency: because the rules are written down, outsiders can inspect what values a model was actually trained toward, instead of inferring them from opaque human feedback data. RLHF and Constitutional AI are typically combined in practice, not treated as either/or alternatives.