Constitutional AI

Constitutional AI treats behavior as something that can be grounded in explicit, human-readable principles — a “constitution” — rather than inferred solely from individual human ratings. The model is trained to ask itself whether a given response violates any of these principles, then revise if it does. This makes the training process more transparent and consistent: instead of hoping that thousands of individual rater decisions happen to encode the right values, you can write down what you believe and test whether the model is actually following it. For behavior architects, it’s a concrete example of how behavioral intent can be operationalized through documented principles rather than vibes.