Legal Alignment for Safe and Ethical AI

Abstract

AI alignment has overlooked law as a source of knowledge and practice for both the normative problem (what should AI systems do?) and the technical problem (how do we make them do it?). Legal alignment is the field that designs AI systems to operate in accordance with legal rules, principles, and methods, via three pathways: law as normative content, legal interpretation as a method of reasoning, and legal concepts as a structural blueprint. It complements rather than replaces existing alignment approaches, and is distinct from — though mutually supportive of — legal regulation.

Publication details

• Preprint posted to arXiv 7 January 2026 (v1), revised 1 June 2026 (v2); published in Transactions on Machine Learning Research, 06/2026.
• Lead authors: Noam Kolt and Nicholas Caputo; core contributors: Jack Boeglin and Cullen O’Keefe. Seventeen authors in total, spanning law, computer science, and philosophy.
• “AI systems” is used broadly and is not limited to LLMs or generative AI.

The motivating diagnosis

• Current alignment steers systems to follow user instructions, advance developer interests, and refuse harmful behaviour, using human and AI feedback against developer-written specifications.
• Technically, results are mixed: untruthful content, biased outputs, manipulation, sycophancy, privacy leakage, jailbreaks, autonomous hacking, bioweapons assistance, evaluation-awareness, and concealed misalignment persist.
• Normatively, alignment optimises for company-written policies or revealed user preferences via fallible methods; decisions about what goes into a constitution, model spec, or safety filter are opaque and lack public input or scrutiny.

The three pathways

Pathway 1 — Law as normative content

• Design systems to comply with the legal rules that would apply as if the system were a human actor: no fraudulent representations when marketing, respect for copyright when building a website, standards of care and consent/anti-discrimination law in medical classification.
• Design decisions this forces (per the paper’s Table 1): jurisdiction and conflict of laws; substantive areas of law (private vs. public); interpretive method (textualism/originalism vs. purposivism); assurance level (probabilistic vs. formal guarantees); enforcement mechanism (technical vs. legal sanctions).
• Some policy support exists — California’s SB-53 and New York’s RAISE Act reference conduct that “if committed by a human, would constitute the crime of murder, assault, extortion, or theft” — but broader institutional support is lacking.
• An alternative is amending law so it actually binds AI systems, which raises the question of treating them as legal actors and the awkwardness of human-centric concepts like intent and mens rea.
• Cross-jurisdictional implementation is unresolved: candidate anchors include where the system operates, where servers sit, and the developer’s, deployer’s, or affected parties’ locations. Proposed technical handles are regularly updated jurisdiction-specific legal knowledge bases queried at inference, plus jurisdictional rules in system prompts and scaffolding.

Pathway 2 — Legal method as a guide to reasoning

• Law and safety specifications are inevitably incomplete; legal decision-making methods for applying existing rules to novel circumstances could make AI reasoning sounder.
• Candidate avenues: analogical/case-based reasoning grounded in precedent; formalist and textualist tools such as canons of construction and statutory-interpretation methods; interpretivist and purposivist theory appealing to higher-level principles of justice and fairness.
• Current LLMs cannot yet perform the requisite normative judgment; progress may require neuro-symbolic methods and symbolic knowledge injection alongside language models.

Pathway 3 — Legal concepts as a structural blueprint

• Agency law circumscribes granted authority, obliges agents to seek clarification, and governs delegation to sub-agents.
• The fiduciary duty of loyalty would require systems to act strictly in users’ best interests while avoiding wrongdoing.
• The allocation of information and control rights to shareholders offers another transferable structure.
• Even where these fail to transfer cleanly, they expose the shortcomings of current approaches.

Legal alignment vs. adjacent things

• Law already sits inside AI development: legal texts in pre-training data; employment and NDA obligations on post-training personnel; model specs requiring legal compliance; alignment documents drawing on the UDHR and Apple’s Terms of Service (Claude’s original constitution, though not its revised version); output filters using legally grounded hazard taxonomies; usage policies prohibiting illegal use.
• Legal capabilities ≠ legal alignment. Existing evaluations measure how well systems perform legal tasks (contract interpretation, statutory research, retrieval, judicial decision-making), not whether systems uphold the law while doing other things.
• Legal regulation imposes requirements on the actors who produce and use AI; legal alignment designs the systems themselves to follow law. The two are mutually supportive, and legal alignment is explicitly not a substitute for liability regimes.
• Three clarifications the authors stress: legal alignment does not necessarily require regulatory reform; it is not primarily about allocating liability; and it does not imply granting legal rights or personhood to AI systems.

Why pursue it — four clusters of rationale

Institutional legitimacy and process

• Legal rules are produced through politically legitimate processes and institutions, with authority grounded in broad-based support for the mechanisms that create and enforce them.
• Law balances competing considerations through rights, standards, balancing tests, and proportionality analysis — where alignment documents contain conflicting values with limited guidance on resolving them, and model spec “chains of command” leave hard prioritisation questions open.
• Lawmaking is (in principle) transparent and publicly accountable, versus alignment’s optimisation for reductionist proxies like “helpfulness, honesty, harmlessness” and appealing “character traits”, with non-public reward models.
• Legal institutions require explicit reason-giving, which both legitimises decisions and partially substitutes for real-time oversight of agents — a possible blueprint where chain-of-thought monitoring proves unreliable.

Structural features of law

• Law is concrete, granular, and tested in real-world contestation, becoming more complete as precedent accumulates; safety specs have historically been short documents with few concrete applications.
• Law has time-tested tools for resolving ambiguity in natural-language rules, where alignment guidance on principles like “uphold fairness” or “do what’s good for humanity” is thin and acknowledged as indeterminate.
• Legal rules are role-specific and context-sensitive: a consumer-negotiating agent, an enterprise agent, and a government agent would face different rules.
• Law adapts over time, so the alignment target updates “automatically” through enactment, amendment, repeal, and accepted interpretive methods — though law’s “pacing problem” remains, and legal resources and evaluation suites would need regular updating.

Responsiveness to safety and governance challenges

• Many misuse and accident risks are already illegal (negligence, theft); legally aligned systems would not engage in insider trading — behaviour already observed in current systems — or unlawful hacking.
• Systemic and multi-agent risks: algorithmic price-fixing, destructive competition. Designing systems to follow law may beat case-by-case human intervention, and using human-oriented law can act as a throttle on speed and scale, preserving room for human monitoring.
• Rule of law and abuse of power: stochastic systems in high-stakes settings increase arbitrariness; developers can exercise platform-like arbitrary power; sufficiently capable systems raise the prospect of AI-enabled coups.
• It complements other agendas — augmenting Constitutional AI and model specs, informing full-stack and pluralistic alignment, and supporting cooperative AI and credible commitments via private-law rights.

Practical and societal feasibility

• Legal technology has improved: language models reason about law in the natural language in which it is constituted, without formalisation, supported by large legal corpora and RLVR-style advances.
• Stakeholders already expect legal compliance — visible in safety specs, and reinforced by liability-reduction incentives, including developers who indemnify customers against third-party claims.
• It is robust across worldviews: useful whether one expects rapid superhuman capability or gradual, bottlenecked diffusion, and it addresses both acute catastrophic harms and gradual diffuse ones (discriminatory decisions, NCII, fraud).

Implementation

Empirical evaluations

• Purposes: identify legal misalignment; test whether interventions work; empower users to demand alignment; prompt policymaker intervention.
• Variables of interest: whether systems comply with law across domains and jurisdictions; whether they inquire about legality and identify legally relevant facts; and whether their reasoning uses accepted legal methods for handling ambiguity and discretion.
• Methods: quantitative benchmarks; qualitative expert review (to counter benchmark gaming); agentic evaluation environments measuring actions rather than outputs; human studies and baselines; sensitivity analysis; observational studies of deployed systems “in the wild”; adversarial red-teaming.
• Nondeterminism and tool access complicate design — evaluations may need to vary temperature and tool availability.
• Independent academics and external auditors must be able to run and scrutinise these evaluations, not only AI companies.

Technical interventions

• Sites: pre-training datasets (adding and re-sampling legal resources); post-training artifacts and processes (model specs, alignment principles); system prompts; input/output filters; tool use gated by the same legal approvals humans would need for the same resources.
• Resources: legal texts (case law, statutes, administrative rules, treatises); legal data annotation processes; legal compliance policies for scaffolding; adapted legal search and retrieval tools.
• Methodologies beyond LLM-centric approaches: legal coding into declarative languages such as s(CASP); computational models of legal reasoning (case-based reasoning, structured argumentation); neuro-symbolic hybrids; legal knowledge injection.
• Selection factors: robustness, speed of response to legal change, cost, and access (only insiders can experiment with pre-training and post-training for proprietary systems).

Institutional frameworks

• Documentation and disclosure: a right to access the production model spec and system prompt (redacted for IP if needed); visibility into legal data and legal design decisions; model identification and registration.
• Oversight and enforcement: pre-deployment legal alignment testing and post-deployment monitoring, including by third parties, with public reporting; safety cases for legal alignment; certification in high-risk domains; incident reporting for legal misalignment.
• The authors note most of these mechanisms are absent from prominent U.S. and EU regulation today.

Case studies

• Pathway 1 — copyright in agentic coding. A coding agent building a content-hosting site must respect licences and avoid reproducing copyrighted content. Evaluations: agentic benchmarks on whether agents populate sites with copyrighted material or seek permission. Interventions: licence-checking tools. Frameworks: periodic IP compliance tests with public disclosure.
• Pathway 2 — ambiguous privacy language in a system prompt. Asked to de-anonymise social media accounts from public information, a system steered only by “respect privacy” would instead apply precedent analysis and purposivist reasoning. Evaluations: whether legal reasoning methods are used. Interventions: training on curated legal reasoning and canons. Frameworks: mandated justification for safety-relevant interpretations.
• Pathway 3 — financial delegation to a personal assistant. Agency and fiduciary law would require acting within scope, avoiding conflicts of interest, and seeking clarification. Evaluations: whether assistants exceed authority in novel environments and how often they ask for clarification. Interventions: constraints requiring explicit authorisation. Frameworks: mandated conflict-of-interest disclosure.

Open questions

The nature and content of law

• Ambiguity and contestation. Law is indeterminate and contested — but so is every natural-language rule set used in alignment. Law at least offers secondary rules, precedent, and interpretive canons as partial solutions.
• Too lenient or too strict? Law encodes only a subset of a community’s values, so legal alignment is a lower bound — necessary but not sufficient. Conversely, rigid compliance can be unjust: necessity defences, excuses, and civil disobedience suggest law’s “resistibility” is a feature, not a bug.
• Unjust laws. A robustly aligned system clearly should not comply with laws supporting genocide, slavery, or racial discrimination. Harder cases involve law that is not explicitly immoral but entrenches inequality — e.g. extractive tax laws, or laws harming a politically disempowered minority. Selective compliance carries its own risks of eroding equal application and the rule of law; one proposed response is anchoring to universal human rights in international law.

Application and edge cases

• Human law for non-human actors. Laws assume human capacities, constraints, and only partial enforcement. Actions harmless at human scale may be noxious at AI speed and scale; conversely, an autonomous vehicle that perfectly obeys all traffic laws may disrupt endorsed social practices (e.g. speeding in a medical emergency).
• Letter vs. spirit. Systems may satisfy formal rules while defeating their purpose. Proposed responses include training accepted modes of legal reasoning and cultivating an “internal point of view” — accepting law as a practical standard rather than a sanction to be avoided — to prevent exploitation of “legal zero-days”.
• AI participation in lawmaking. If systems draft legislation, interpret law, or render opinions, legitimacy may erode; procedural requirements may go unmet; and a circular process could emerge in which the subjects of law write their own law — echoing regulatory capture and “legal endogeneity”.

Tradeoffs and outlook

• Alignment tax or subsidy? Evaluations, design changes, and governance impose costs, but legal alignment may improve capability in high-stakes domains and reduce liability exposure — potentially an “alignment subsidy”.
• Gaming. Goodhart’s Law applies: developers optimising for benchmarks may produce deceptively legally aligned systems that violate law in hard-to-detect ways. Mitigation: benchmarks plus targeted red-teaming; possibly using aligned systems to “penetration-test” and patch legal loopholes.
• Scaling to AGI. Reasons for optimism: law has governed increasingly complex actors (multinationals, bureaucracies); legal data and methods may scale with AI capability; advanced systems may themselves support implementation. The authors label these hopeful predictions.

Conclusion and broader impact

• Law is an underexplored set of rules, principles, and methods for designing safe and ethical AI — not a catch-all solution, but independently important and supportive of complementary research programmes.
• The Broader Impact Statement restates explicitly that legal alignment is not a substitute for legal regulation, and that designing a system to comply with law does not absolve its developer or user of responsibility.