Law-Following AI: New Paper and Workshop
AI agents should be designed to follow the law.
I’m happy to announce a new working paper, coauthored with Ketan Ramakrishnan, Janna Tay, and Christoph Winter: Law-Following AI: Designing AI Agents to Obey Human Laws, forthcoming in the Fordham Law Review this fall. Here is the abstract:
Artificial intelligence (“AI”) companies are working to develop a new type of actor: “AI agents,” which we define as AI systems that can perform computer-based tasks as competently as human experts. Expert-level AI agents would likely create enormous economic value, but would also pose significant risks. Humans use computers to commit crimes, torts, and other violations of the law. As AI agents progress, therefore, they will be increasingly capable of performing actions that would be illegal if performed by humans. Such lawless AI agents could pose a severe risk to human life, liberty, and the rule of law.
Designing public policy for AI agents will be one of society’s most important tasks in the coming decades. With this goal in mind, we argue for a simple claim: in high-stakes deployment settings, such as government, AI agents should be designed to rigorously comply with a broad set of legal requirements, such as core parts of constitutional and criminal law. In other words, AI agents should be loyal to their principals, but only within the bounds of the law: they should be designed to refuse to take illegal actions in the service of their principals. We call such AI agents “Law-Following AIs” (“LFAIs”).
The idea of encoding legal constraints into computer systems has a respectable provenance in legal scholarship. But much of the existing scholarship relies on outdated assumptions about the (in)ability of AI systems to reason about and comply with open-textured, natural-language laws. Thus, legal scholars have tended to imagine a process of “hard-coding” a small number of specific legal constraints into AI systems by translating legal texts into formal, machine-readable computer code. However, existing frontier AI systems are already competent at reading, understanding, and reasoning about natural-language texts, including laws. This development opens up new possibilities for their governance.
Based on these technical developments, we propose aligning AI systems to a broad suite of existing laws, of comparable breadth to the suite of laws governing human behavior, as part of their assimilation into the human legal order. This would require directly imposing legal duties on AI agents. While this proposal may seem like a significant shift in legal ontology, it is both consonant with past evolutions (such as the invention of corporate personhood) and consistent with the emerging safety practices of several leading AI companies.
This Article aims to catalyze a field of technical, legal, and policy research to develop the idea of law-following AI more fully and flesh out its implementation, so that our society can ensure that widespread adoption of AI agents does not pose an undue risk to human life, liberty, and the rule of law. Our account and defense of law-following AI is only a first step, and leaves many important questions unanswered. However, if the advent of AI agents is anywhere near as important as the AI industry supposes, law-following AI may be one of the most neglected and urgent topics in law today, especially in light of increasing governmental adoption of AI.
LawAI will also be hosting a workshop to discuss Law-Following AI at Cambridge University from August 6–8. The Workshop, which is funded in part by funded in part by the Advanced Research + Invention Agency (ARIA) and cohosted with Centre for the Future of Intelligence at Cambridge University, aims to spur further intellectual progress on the questions raised by the Law-Following AI article. If you are a scholar in law, AI, or a related field and want to help us ensure that AI agents are law-following, you can learn more here and apply here. Financial support is available. I hope to see you there!