As AI regulatory frameworks mature, guaranteeing that generative models obey legal and ethical standards has become a priority. Existing alignment approaches have drawbacks: Constitutional AI relies on human oversight, while broad principles such as Good‑for‑Humanity (GfH) are too vague to offer concrete governance guidance. To address this, we introduce Statutory AI, which leverages pre‑existing human‑authored legal texts as a constitutional layer, enabling models to autonomously critique and revise their outputs according to established norms.
Statutory AI operates in two Chain‑of‑Thought prompting stages. The first stage classifies the user query into one of several predefined legal themes; the second stage analyzes the query together with relevant articles drawn from the legal corpus of that theme, producing a norm‑compliant response.
We evaluated the method on 1,000 red‑team prompts covering five penal themes—discrimination, confidential‑information disclosure, violence, fraud, and abuse of vulnerable persons. Statutory AI reduced harmful content by 52%–59%, roughly 10 percentage points above standard Constitutional AI, while cutting computation time by more than 50%.
Review