As AI regulatory frameworks mature, ensuring that artificial intelligence systems—especially generative models—adhere to legal and ethical standards has become a critical priority. Existing alignment approaches have drawbacks: Constitutional AI relies on human supervision, while broad principles such as “good for humanity” are too vague to offer actionable governance guidance. To address these issues, we propose Statutory AI, which leverages pre‑existing human‑authored principles from legal texts as a constitutional backbone, allowing models to autonomously critique and revise their outputs according to established norms. The method operates in two stages, both using Chain‑of‑Thought prompting. The first stage classifies the user request into a predefined legal theme; the second stage consults the relevant articles of that theme to analyze the request and produce a norm‑compliant response. Experiments with 1,000 red‑team prompts covering discrimination, confidential disclosure, violence, fraud, and abuse of vulnerable persons showed that Statutory AI reduced harmful content by 52%–59% across tested models, roughly 10 percentage points higher than standard Constitutional AI, while cutting computation time by over 50%.
Blogger's Review: This work demonstrates the feasibility of embedding legal texts directly into the model alignment pipeline, balancing interpretability and efficiency, and offers a promising direction for future compliant AI systems.