Red Teaming the Law: An Adversarial Approach to Legal Alignment

A core steering mechanism in a piece of regulation may lie not in what is written within it --- the primal path it lays to follow its rules--- but in what is left out of it --- the dual path it lays to avoid them. In this early-stage work, we frame a process of legal alignment (the alignment of a piece of regulation's text with its goals) as a multi-agent zero-sum game. Our hope is to demonstrate the relevance of AI alignment methods to legal oversight and the value of an adversarial frame for policy design.