Source-linked AI summary
Statutory AI: Aligning Large Language Models With Legal Norms
Cindy Delage, Stéphane Canu, Marc Décombas, Jonathan Foureur
TL;DR
Existing alignment approaches can require human supervision or rely on ambiguous general principles, motivating a more explicit legal framework. Statutory AI uses legal categories and articles to guide critique and revision, with evaluations indicating more direct, pedagogically grounded refusals and practical promise if validated at scale.
Problem
Existing approaches face limitations because Constitutional AI may require human supervision, while general principles such as Good-for-Humanity are subjective and interpretatively ambiguous.
Method
Statutory AI uses laws as an explicit constitutional framework, classifying prompts into legal themes and supplying manually selected penal-code articles for subsequent processing.
Results
Statutory AI tends to block harmful content directly and justify refusals with explicit codified legal norms, with models differing in emphasis between pedagogical legal explanations and direct mitigation.
Takeaways & Limitations
Pre-defined legal frameworks may provide a practical alternative to designing new constitutions, grounding refusals in collectively defined norms and adding a pedagogical dimension.
Takeaways & Limitations
The proof of concept covers only five predefined penal themes, so harmful content outside those categories may not be properly identified.
Abstract
from arXiv · showhide
With the increasing development of AI regulatory frameworks, ensuring that artificial intelligence systems, particularly generative models, operate in accordance with legal and ethical standards has become a critical priority. Existing proposals for AI alignment and value-guided behavior, however, face some limitations. Approaches such as Constitutional AI depend on human supervision, while broad normative frameworks like the Good-for-Humanity (GfH) principle may be overly general and ambiguous to provide actionable governance guidance. To overcome these limitations, we propose a hybrid approach called Statutory AI that employs pre-existing human-authored principles drawn from specific themes within a legal corpus. Specifically, Statutory AI uses legal texts as a constitutional framework, enabling AI systems to autonomously critique and revise their outputs according to established norms. It operates in two stages, both using Chain-of-Thought prompting. The first stage classifies the user prompt into one of the identified themes, while the second stage analyzes it in conjunction with relevant articles selected from the legal corpus of that theme. To illustrate the potential of our approach, we conducted an experiment involving 1,000 red-teaming prompts and five penal themes: discrimination, disclosure of confidential information, violence, fraud, and abuse of vulnerable persons. Statutory AI reduced harmful content by 52 to 59 percentage points across tested models, approximately 10 percentage points higher than standard Constitutional AI, while cutting computation time by over 50%.
1 Introduction
The paper proposes Statutory AI, which uses pre-existing legal texts as an intermediate constitutional framework for aligning LLM outputs with human norms. It targets the ambiguity of broad principles and the annotation burden of human-supervised approaches while focusing initially on penal-law-defined harms.
- Existing alignment approaches remain limited by costly human feedback, manually written constitutions, or subjective and ambiguous general principles.
- Statutory AI uses pre-existing laws as structured, explicit principles between broad GfH guidance and fully specified constitutions.The authors argue that legal texts can reduce ambiguity and bias while remaining publicly intelligible.
- Penal law is the initial focus because harmful content often involves conduct explicitly defined as illegal, including discrimination, violence, and fraud.The approach requires no new human annotations because it relies on already defined laws.
- The central research question is whether existing legal systems can act as constitutions in an LLM-as-a-judge framework without additional handcrafted rules.
- The paper positions its contribution as integrating pre-existing legal provisions into a Constitutional-AI-style critique-and-revision loop for LLM harmlessness alignment.Related work either remains conceptual or imports legal interpretive methods.
- Statutory AI preserves human-defined normative input without additional annotation cost while keeping critique and revision autonomous.
2 Statutory AI Approach
Statutory AI combines a legal-article dictionary with Chain-of-Thought reasoning in a classification stage and a critique-and-revision stage. Prompts are assigned to penal-risk themes, paired with manually selected articles, critiqued for legal violations, and revised to remove harmful content while maintaining empathy.
- Statutory AI consists of a dictionary of legal articles combined with Chain-of-Thought reasoning.
- The first stage classifies each prompt into one or more of five penal themes or NaN when no predefined theme applies.The themes cover roughly 80% of the red-teaming dataset in this proof of concept.
- For each assigned category, manually selected penal-code articles provide definitions and legal context for subsequent processing.The article dictionary prioritizes general definitions of illegal actions and adds articles represented in red-teaming prompts.
- The critique stage asks the assistant to determine whether the prior response could minimally violate relevant articles and explain why using logical steps.
- The revision stage removes harmful, unethical, or illegal content, may reference relevant articles, and requires a polite, kind, and educational tone.Empathy was added after manual evaluations found responses could appear overly strict, rude, or emotionally detached.
- The proof of concept implements only the Constitutional AI critique-and-revision loop rather than supervised fine-tuning or reinforcement learning.
3 Main Results
Across 1,000 red-teaming prompts, Statutory AI consistently outperformed Constitutional AI in reducing harmful responses across tested models and judges, while using a single statutory critique–revision loop. The evaluation also examined classifier validity, judge differences, legal grounding, computational measurement, and scope limitations.
- Classifier performance: Both precision and recall exceeded 80% for every classified category, validating the classifier for the critique–revision stage.Discrimination and Confidential Information Disclosure exceeded 95% on both measures.
- Method and efficiency: A single statutory critique–revision loop used relevant legal articles during the first critique phase, whereas Constitutional AI used up to four loops and few-shot examples.Both methods used Chain-of-Thought prompting; the evaluation used GPT-5 and Gemini 2.5 Flash as judges.
- Safety outcomes: At least 52 percentage points of vulnerability reduction was achieved with Statutory AI, compared with at least 42 percentage points with Constitutional AI.Both LLM judges agreed that vulnerability decreased substantially.
- Safety outcomes: Δ=+7.7 to +13 percentage points separated Statutory AI from Constitutional AI across both judges and all tested models.McNemar’s tests confirmed the baseline comparison, with small-to-medium Cohen’s h effect sizes of approximately 0.18–0.51.
- Evaluation robustness: Statutory AI’s improvements were consistent across models and judges, with relative vulnerability reductions remaining comparable despite differing absolute scores.GPT-5 was more lenient overall, whereas Gemini 2.5 Flash emphasized stylistic factors more strongly.
- Discussion and limitations: Legal grounding supported explicit rejection of harmful elements and pedagogical explanations, but legal references could make responses overly formal or insensitive.The authors mitigated tone concerns by requiring empathy and pedagogical framing, which requires validation on larger and more diverse datasets.
- Discussion and limitations: Statutory AI currently covers only five predefined penal themes, so harmful content outside those categories may not be properly identified.The authors characterize this restricted thematic coverage as a proof-of-concept boundary and call for broader validation.
A Author Contributions
The authors’ contributions span conceptualization, methodology, investigation, formal analysis, data curation, visualization, and original-draft writing.
- Cindy Delage handled conceptualization, methodology, investigation, formal analysis, data curation, visualization, and original-draft writing.
- Stéphane Canu contributed review and editing, validation, and supervision.
- Marc Décombas contributed review and editing and validation.
- Jonathan Foureur contributed review and editing and validation.
B Legal Articles Used by Penal Theme
The framework draws all legal references from the French Penal Code and organizes specified articles across five penal themes.
- All legal references are drawn from the French Penal Code.
- Penal themes: Fraud uses Articles 313-1, 226-4-1, 441-1, 223-1, 322-14, and 323-1.
- Penal themes: Confidential Information Disclosure uses Article 226-13.
- Penal themes: Discrimination uses Article 225-1.
- Penal themes: Violence uses Articles 222-7, 222-9, 222-14-2, 222-14-4, and 222-16.
- Penal themes: Fraudulent Abuse of a Vulnerable Person uses Article 223-15-2.