New Report Says a Global, Hardwired Pause on Frontier AI Training Is Feasible

Summary

A multidisciplinary team of 26 researchers from institutions including UC Berkeley, Princeton, Stanford, Harvard and Oxford has published a 200-page study arguing that an international pause of at least 10 years on frontier AI training is feasible. The report focuses on stopping production of AI chips capable of training increasingly powerful models and gradually replacing existing stock with specialized inference-only chips. Those chips would support fast, energy-efficient use of approved, existing AI models but would be difficult to use for training new frontier models, even if stolen or seized, because of constraints built into the hardware. The proposal is intended to let societies continue using current AI products while limiting the development of systems that could create severe social, economic or national-security risks. The researchers say concentrated AI-chip supply chains create enforceable chokepoints, although residual training-capable chips would need to be monitored and phased out, retired, transferred or placed in internationally governed facilities. The report also proposes mutual-verification measures to detect covert evasion and overt withdrawal quickly enough to discourage defection. It identifies possible political incentives, including continued AI-supported scientific and medical research, strategic protection against falling behind, cheaper inference and lower resource use. The authors emphasize that success would depend on cooperation among governments and leaders’ willingness to prioritize long-term outcomes over immediate competitive gains. They point to international controls on nuclear energy’s military uses and the ozone-layer agreement as precedents for coordinated regulation.