UC Berkeley Professors Propose Multilateral Governance Framework for Global Pause on Frontier AI

A prominent group of computer science professors and public policy scholars at UC Berkeley released a comprehensive policy blueprint on Wednesday, October 21, proposing an enforceable international treaty to pause the development of unaligned frontier artificial intelligence systems exceeding specified compute thresholds. Published through the Berkeley Center for Long-Term Cybersecurity, the proposal outlines diplomatic and technological mechanisms designed to monitor global semiconductor supply chains and verify compliance without compromising corporate intellectual property.

The framework recommends establishing a multilateral AI Safety Directorate modeled after the International Atomic Energy Agency, empowered to maintain cryptographic telemetry registries on advanced lithography equipment and high-bandwidth memory production. Under the treaty terms, commercial and state laboratories would be prohibited from initiating training runs requiring more than 10^26 floating-point operations until independent safety auditors and international commission officials certify that automated systems cannot autonomously replicate, deceive human overseers, or execute cyber warfare operations.

Co-authors emphasized that existing voluntary industry commitments and fragmented national executive orders are insufficient to mitigate catastrophic risks stemming from competitive race dynamics between leading frontier laboratories in San Francisco and international rivals. Department researchers argued that hardware-level compute governance represents the only verifiable choke point for enforcing safety benchmarks across distributed sovereign cloud networks.

Academic institutions, technology ethics boards, and congressional committee staff will review the Berkeley proposal during an open symposium scheduled for Friday at Sproul Hall in Berkeley. The complete 45-page policy document, including draft treaty protocols and hardware verification whitepapers, is accessible online at cltc.berkeley.edu. Policy inquiries may be directed to the Berkeley Center for Long-Term Cybersecurity at (510) 642-1000.