When Anthropic published its Constitutional AI paper in late 2022, the alignment research community received it with a mixture of genuine intellectual interest and cautious scepticism. The proposal — that an AI system could be trained to evaluate and revise its own outputs against a set of explicit principles, reducing reliance on human feedback for harmful content identification — was elegant in conception and uncertain in practice. Two years of deployment at scale has provided enough evidence to move from speculation to provisional assessment.

The positive findings are meaningful. Constitutional AI has demonstrably reduced the rate of harmful output in the categories it targets: content that is directly dangerous, clearly deceptive, or in violation of basic ethical norms. The models trained with constitutional methods require fewer examples of problematic content in their training data, reducing the burden on human annotators who must otherwise review disturbing material at scale. The approach has also proven more amenable to targeted refinement than earlier reinforcement learning from human feedback methods, allowing the company to update the model's behaviour in specific domains without extensive retraining.

The limitations are equally real. Constitutional AI is effective at addressing the content categories its constitution addresses; it is less effective at the subtler forms of problematic behaviour that emerge from capable models operating in complex social and institutional contexts.