What would multilateral ‘AI arms control’ look like?

Existing international arms control treaties, especially those related to nuclear weapons, offer a valuable starting point for developing a multilateral framework for AI arms control. The historical record indicates that such treaties can provide a blueprint for how to handle the complexities of emerging technologies, setting precedents for monitoring, verification, and international cooperation. However, the dual-use nature of AI, where technologies can have both beneficial and harmful applications, presents a significant challenge that was less prevalent with materials like sarin, which has no civilian use.

A key aspect of any effective AI treaty would be a robust inspection regime. Historical data from arms control agreements, such as those monitored by the IAEA for nuclear weapons and the OPCW for chemical weapons, suggest a hybrid model. This model would likely involve continuous monitoring as a primary layer, augmented by on-site inspections triggered by anomalies or state-party requests. Unlike the bilateral, party-on-party inspections common in some historical treaties, a multilateral AI agreement would necessitate an independent institutional inspectorate, similar to the OPCW, to manage the complexities of monitoring numerous participants and ensure findings are not dismissed as politically motivated.

The proposed inspection architecture for AI could feature a three-layer system. The primary layer would be a Global AI Monitoring Network, analyzing power-grid data, semiconductor exports, cloud API patterns, and voluntary telemetry. A secondary layer would involve on-site inspections for anomalies or routine visits to top-tier facilities. A critical innovation would be an “automatic compute-anomaly trigger” that bypasses the need for state accusation, addressing the historical issue where formal challenge inspections often go uninvoked due to geopolitical considerations. Furthermore, lessons from talent retention challenges in organizations like the OPCW, where inspectors earn significantly less than top-tier AI engineers, highlight the need for competitive compensation to attract and retain expert AI inspectors. The rapid evolution of AI capabilities, on cycles of 6-24 months, also demands that inspection protocols be continuously revised, unlike the more stable nature of chemical weapon protocols.