Abstract
Advanced AI agents raise significant challenges to global stability, cooperation, and international law. States increasingly face pressure to manage the geopolitical, safety and security risks such agents give rise to. Yet advanced AI agreements for these systems face significant hurdles, including potentially unpalatable compliance monitoring requirements, and the ease with which states or their ‘lawless’ AI agents might exploit legal loopholes. This article introduces a novel commitment mechanism for states: Treaty-Following AI (TFAI). Building on recent work on “Law-Following AI”, we propose that AI agents could be technically and legally designed to execute their principals' instructions except where those involve goals or actions that would breach a designated AI-Guiding Treaty. If technically and legally feasible, the TFAI framework offers a powerful, verifiable and self-executing commitment mechanism by which states could guarantee that their deployed AI agents will abide by the legal obligations to which those states have agreed. This could unlock otherwise-unfeasible advanced AI agreements, and potentially help revitalize treaties as an instrument across many domains of international law. The article examines the conceptual foundations, technical feasibility, political uses, and legal construction of TFAI agents and AI-guiding treaties, answering questions around the design of AI-guiding treaties, state responsibility for TFAI agents, and appropriate methods of treaty interpretation for TFAI agents, amongst others. It argues that, if these outstanding technical, legal, and political hurdles are cleared, treaty-following AI could serve as an important scalable cooperative capability for both the international governance of advanced AI, and for the broader integrity of international law in the ‘intelligence age’.
Matthijs Maas
Senior Research Fellow
I am a Senior Research Fellow at the Institute for Law & AI, working on adaptive global governance approaches for AI.