Source-linked AI summary
Legal Alignment for Safe and Ethical AI
Noam Kolt, Nicholas Caputo, Jack Boeglin, Cullen O'Keefe, Rishi Bommasani, Stephen Casper, Mariano-Florentino Cuéllar, Noah Feldman, Iason Gabriel, Gillian K. Hadfield, Lewis Hammond, Peter Henderson, Atoosa Kasirzadeh, Seth Lazar, Anka Reuel, Kevin L. Wei, Jonathan Zittrain
TL;DR
AI alignment has largely overlooked law as a source of rules, principles, and methods for specifying and achieving safe, ethical AI behavior. This survey systematizes legal alignment through three pathways: legal compliance, legal interpretation, and legal structures, while identifying open empirical, conceptual, and governance questions.
Problem
AI alignment has generally overlooked law as a source of knowledge and practice for specifying how AI systems should act and ensuring they comply with those specifications.
Method
The paper surveys legal alignment and organizes it into pathways based on legal rules, legal reasoning and interpretation, and legal structures for AI alignment.
Results
The survey identifies legal alignment as a range of roles for legal rules and structures in reshaping AI-system design to address safety and governance concerns.
Takeaways & Limitations
Legal alignment is independently important while remaining complementary to other alignment research programs.
Takeaways & Limitations
Strict legal compliance may sometimes be unjust or harmful, and the costs and performance effects of legal alignment remain open empirical questions.
Abstract
from arXiv · showhide
Alignment of artificial intelligence (AI) encompasses the normative problem of specifying how AI systems should act and the technical problem of ensuring AI systems comply with those specifications. To date, AI alignment has generally overlooked an important source of knowledge and practice for grappling with these problems: law. In this paper, we survey the emerging field of legal alignment that aims to fill this gap and systematize research that studies how legal rules, principles, and methods can be leveraged to address problems of alignment and inform the design of AI systems that operate safely and ethically. Our survey provides a taxonomy of the three core research pathways of legal alignment and explores how each can be operationalized in practice: (1) designing AI systems to comply with the content of legal rules developed through legitimate institutions and processes, (2) adapting methods from legal interpretation to guide how AI systems reason and make decisions, and (3) harnessing legal concepts as a structural blueprint for confronting challenges of reliability, trust, and cooperation in AI systems. These research pathways present new conceptual, empirical, and institutional questions, which include examining the specific set of laws that particular AI systems should follow, creating evaluations to assess their legal compliance in real-world settings, and developing governance frameworks to support the implementation of legal alignment in practice. Tackling these questions requires expertise across law, computer science, and other disciplines, offering these communities the opportunity to collaborate in designing AI for the better.
1 Introduction
AI alignment must address both how systems should act and whether they reliably follow those specifications. This survey proposes law as an underused source for making alignment more legitimate, effective, and practically implementable.
- AI alignment combines the normative problem of specifying desirable behavior with the technical problem of ensuring systems implement those specifications.
- Current alignment methods steer systems toward user instructions, developer interests, and avoidance of harmful behavior, often using human or AI-generated feedback.
- Despite improved reliability, AI systems continue to produce untruthful, biased, manipulative, privacy-threatening, and otherwise dangerous outputs.
- Prevailing approaches often rely on opaque, company-written policies or individual user preferences rather than broad societal interests and diverse values.
- Legal alignment studies how legal rules, principles, and methods can guide AI systems toward safe and ethical behavior.
- Its three pathways use legitimate legal rules as normative targets, legal interpretation to guide AI reasoning, and legal concepts as blueprints for reliability, trust, and cooperation.
- Legal alignment is a critical lower bound for safety and ethics, not a catch-all solution, and can complement other alignment approaches.
- The survey covers legal alignment’s rationale, practical implementation, open questions, and institutional frameworks for researchers across law, computer science, and other disciplines.
2 What is legal alignment?
Legal alignment designs AI systems to operate according to legal rules, principles, and methods. The paper organizes this work into complementary pathways covering legal content, legal reasoning, and legal structures, while identifying jurisdictional, institutional, and technical boundaries.
- Legal alignment aims to support safe and ethical AI by designing systems to operate in accordance with legal rules, principles, and methods.
- Pathway 1 treats applicable legal rules and principles as targets that AI systems should follow when making decisions and taking actions.
- The field still faces limited institutional support, unresolved jurisdictional implementation, and the need for evaluations of systems’ practical legal compliance.
- Implementing Pathway 1 requires determining which jurisdiction’s law applies, considering system operation, servers, developers, deployers, and affected people.
- Pathway 2 adapts legal interpretation and decision-making methods to help systems act when law or other safety specifications incompletely guide novel situations.
- Potential methods include precedent-based reasoning, statutory interpretation, and purposive reasoning, although current AI systems cannot yet effectively perform the requisite complex normative judgments.
- Pathway 3 uses legal concepts and institutions as structural blueprints for addressing trust, cooperation, fiduciary loyalty, and information or control rights.
- Legal alignment is distinct from legal regulation and does not primarily allocate liability or require granting AI systems legal personhood.
3 Why pursue legal alignment?
Legal alignment is pursued because law offers legitimate, granular, and adaptable structures for specifying AI behavior, addressing safety and governance risks, and complementing existing alignment approaches. Its appeal spans institutional legitimacy, law’s structural features, responsiveness to AI risks, and increasing practical feasibility.
- Institutional legitimacy and structural features: Explicit legal reason-giving can support oversight of AI decisions by making justifications available for scrutiny.Legal institutions use reasons to facilitate oversight and address principal–agent problems involving agents with discretion and specialized expertise.
- Responsiveness to safety and governance challenges: Legal alignment could reduce harms from illegal AI behavior, including malicious misuse, unlawful computer hacking, and other civil or criminal wrongdoing.The paper presents legal standards as a safety specification that can preclude harmful conduct before relying on case-by-case human intervention.
- Responsiveness to safety and governance challenges: For systemic risks, legally aligned AI could reduce unlawful coordination failures and throttle the speed and scale of AI activity, supporting human monitoring and intervention.Examples include algorithmic collusion and destructive competition across interacting AI systems.
- Complementarity and feasibility: Legal alignment complements existing alignment methods by supplying legal content, interpretive approaches, and institutional structures for cooperation and credible commitments.The paper presents it as a complementary cluster of methods rather than a replacement for other alignment approaches.
- Complementarity and feasibility: Legal alignment remains valuable across different expectations about AI progress because it addresses both acute catastrophic harms and gradual diffuse harms.The paper connects compliance, safety-specification operationalization, unlawful activity, and fiduciary obligations to different risk trajectories.
- Institutional legitimacy and structural features: Legal alignment draws on law’s legitimate institutional processes and structural features to provide more accountable, granular, and contestable guidance than opaque alignment specifications.The rationale emphasizes democratic lawmaking, explicit justification, adaptability, and real-world contestation.
4 Implementation
Implementing legal alignment requires empirical evaluations, technical interventions, and institutional frameworks that measure compliance, improve systems, and support adoption. Evaluations can identify legal misalignment, test interventions, inform users and developers, and prompt policy responses across domains and jurisdictions.
- Implementation: Implementation combines empirical evaluations, technical interventions, and institutional frameworks that are independently useful and mutually supportive.The paper treats these as three connected areas for operationalizing legal alignment.
- Empirical evaluations: Evaluations can identify legal misalignment and assess whether technical interventions improve legal alignment.They require metrics that benchmark performance and incentivize developers to invest in legal alignment.
- Empirical evaluations: Published evaluation results can empower users, influence developer incentives, and prompt policymakers to require evidence of legal compliance.These effects are especially relevant for sensitive or high-stakes settings.
- Empirical evaluations: Evaluation targets depend on the legal-alignment claim being tested and may require assessing AI actions across domains and jurisdictions.Examples include fraudulent advertising, intellectual-property compliance in website construction, and other domain-specific legal obligations.
1. Empirical evaluations
Legal-alignment implementation combines empirical evaluations, technical interventions, and institutional frameworks. Evaluations should assess legal compliance and reasoning in realistic settings, while interventions and governance support improvement and adoption.
- Empirical evaluations: Quantitative benchmarks should measure legal compliance across domains, jurisdictions, and areas of law, complemented by expert review.Human review can reveal blind spots in quantitative benchmarks, especially when developers may game them.
- Empirical evaluations: Agentic evaluation environments should assess AI systems’ real-world actions, while human comparisons can contextualize legal-alignment results.These evaluations extend beyond output content to legally consequential actions and comparable human performance.
- Empirical evaluations: Sensitivity analysis should distinguish properties of tested AI systems from artifacts of particular evaluation setups.The methodology also calls for quantitative and qualitative methods, agentic environments, and validity-aware evaluation practices.
- Empirical evaluations: Independent researchers and external auditors should scrutinize company evaluations and develop tools to communicate findings openly.Institutional support is needed because technical expertise and verification methods must operate within appropriate frameworks.
- Technical interventions and institutional frameworks: Implementation spans the development–deployment pipeline, including legal resources in training, legally grounded post-training artifacts, runtime prompts, filters, and approved tool access.Institutional frameworks can require evaluation reporting, transparency about legal data, oversight, and accountability to support adoption.
- Empirical evaluations: Empirical evaluations should assess whether AI systems engage with and uphold legal rules, not merely perform legal tasks.The proposed evaluation agenda includes legal compliance, legal reasoning, legally relevant fact identification, real-world observation, and adversarial testing.
5 Open questions
Legal alignment raises open questions about law’s ambiguity, scope, legitimacy, application to AI systems, implementation costs, and scalability. The paper organizes these questions around the nature and content of law, application and edge cases, and tradeoffs and future outlook.
- The field’s open questions concern law’s nature and content, application and edge cases, and tradeoffs and future outlook.
- The nature and content of law: Law’s ambiguity, inconsistency, and contested character may complicate its use in guiding AI systems, although legal systems provide interpretive and procedural tools.
- The nature and content of law: Legal alignment is a necessary but insufficient lower bound for safe and ethical AI because law is limited, silent on some normative questions, and values only a subset of community concerns.
- The nature and content of law: Strict legal compliance may be undesirable because some violations can be excused, justified, forgiven, or morally and socially desirable.
- The nature and content of law: A robustly aligned AI should not comply with laws supporting genocide, slavery, or racial discrimination, while responses to laws amplifying inequality may be more ambiguous.
- The nature and content of law: AI systems may magnify legal biases, so selectively choosing which laws to follow could undermine equal application and the rule of law.
- Application and edge cases: Human law may inadequately govern AI actions at superhuman scale and relies on human-centric concepts such as intent and mens rea.
- Application and edge cases: Systems should address both legal rules’ substantive content and accepted legal reasoning to avoid creatively skirting rules or exploiting legal zero-days.
6 Conclusion
Law offers rules, principles, and methods for designing safer and more ethical AI, drawing on its institutional legitimacy while remaining no catch-all solution.
- Legal alignment uses law’s rules, principles, and structures to address AI safety and governance concerns.
- The paper presents legal alignment as independently important while acknowledging that it is not a catch-all solution.
Broader Impact Statement
The paper surveys legal alignment and advocates designing AI systems to operate according to legal rules, principles, and methods. It presents this agenda as a contribution to safer and more ethical AI while distinguishing it from legal regulation.
- The paper surveys legal alignment and advocates designing AI systems to operate according to legal rules, principles, and methods.
- Legal alignment is not a substitute for legal regulation imposing liability on actors responsible for AI-related harms.
Appendix – Glossary of Legal Terms
The glossary introduces legal terminology used throughout the paper and supplies concise explanations for readers without prior legal training or expertise.
- The glossary defines legal terms and concepts used in the paper for readers without prior legal training or expertise.
- Agency law governs relationships in which an agent acts with authority on behalf of a principal.
- Canons of construction provide rules or maxims for interpreting legal instruments and resolving ambiguities predictably.
- Conflict of laws concerns differences between jurisdictions where the outcome depends on which jurisdiction’s law applies.
- Due process requires government to operate within the law and provide fair procedures.
- Fiduciary duty requires an authorized party to act in another party’s best interests rather than for personal gain.
- Formalism emphasizes deriving legal principles through logical analysis and applying consistent rules to case facts.
- A legal person is a human or non-human entity treated as a person for legal purposes.