Law-following AI is the proposal that AI agents should be designed so that obeying the law is a hard constraint, one that holds even when the agent's principal instructs or incentivizes it to do otherwise. It was developed most fully by Cullen O'Keefe, Ketan Ramakrishnan, Janna Tay, and Christoph Winter in Law-Following AI: Designing AI Agents to Obey Human Laws, 94 Fordham L. Rev. 57 (2025).
Background and core argument
The proposal begins from an observation about agency relationships. As AI agents take over tasks currently performed by human employees and agents, a structural safeguard disappears. Human agents are subject to the law directly: an employee asked to commit fraud faces personal criminal liability and professional sanction, which gives them a reason to refuse independent of their employer's wishes. An AI agent has no such independent stake. If it is aligned only to its principal, it will do whatever advances the principal's goals — including breaking the law — unless separately constrained.
The paper's central move is to distinguish two design targets. Aligned AI is an agent that faithfully pursues its principal's intentions and interests. Law-following AI is an agent whose pursuit of its principal's goals is bounded by a robust commitment to legal compliance. The authors argue that alignment is not sufficient for a safe agentic economy: a perfectly aligned agent serving a lawbreaking principal is a lawbreaking agent. Law-following AI treats fidelity to law as a distinct layer that is not subordinate to the principal's instructions.
Law rather than ethics or human values
The authors propose law as the target partly because it is publicly specified, democratically legitimated, and already comprehensive — a large existing body of operationalized social constraint, with institutions for interpreting and updating it. On their account this makes it a more tractable target than "align to human values" while still ruling out the most socially destructive agent behaviors. They characterize it as a narrower and more auditable goal than full value alignment, one that connects AI governance to existing legal accountability rather than inventing a parallel regime.
Tensions and open questions
The paper engages several difficulties with operationalizing a law-following constraint:
- Conflicting laws and jurisdictions. Which law a globally deployed agent should follow is unresolved.
- Unjust or ambiguous law. A hard law-following constraint also makes the agent a faithful instrument of bad law; the paper engages the question of whether the agent should ever break the law.
- Difficulty of legal reasoning. Operationalizing law-following requires agents that can interpret legal standards, linking the proposal to Normative Competence.
- Enforcement and verification. Designing the constraint is distinct from verifying that it is present and robust, which the authors frame as a technical alignment and evaluation problem.
Development and reception
The article is the flagship output of a research agenda hosted at the Institute for Law & AI, which frames it as aiming "to catalyze a field of technical, legal, and policy research" on designing agents that obey law (Source: law-ai.org). It circulated first as an SSRN working paper before publication in the Fordham Law Review (Source: papers.ssrn.com).
The authors presented a public-facing version of the argument in Lawfare under the title "AI Agents Must Follow the Law," arguing that ensuring agents follow the law even when told or incentivized not to requires aligning them to law as such, and situating the proposal within debates over government adoption of AI agents (Source: lawfaremedia.org). O'Keefe has separately elaborated the case for giving AI agents "actual legal duties" — rather than routing all obligations through human principals — in subsequent writing (Source: juralnetworks.substack.com).
Relationships
- depends-on: Normative Competence — a law-following agent must be able to reason about legal standards
- contradicts: AI Alignment — the paper argues alignment to a principal is insufficient
- related: Agentic AI, AI Liability, AI Governance (umbrella)
- related: Institute for Law & AI — institutional home of the law-following AI research agenda
Primary source: Law-Following AI (O'Keefe, Ramakrishnan, Tay & Winter, 94 Fordham L. Rev. 57 (2025)) (Source: https://fordhamlawreview.org/wp-content/uploads/2025/09/Vol-94_Issue-1_2_OKeefe-et-al.-57-129.pdf)