Researchers say a 10-year global pause on frontier AI training is feasible

A first-of-its-kind academic study

According to UC Berkeley, the study is the first of its kind in academia. The Working Group on AI Pause Feasibility brings together faculty from UC Berkeley, Princeton, Stanford, Harvard and Oxford, among other institutions. Their fields include computer science, economics, history, international relations, and Chinese and American politics, and some members have practical experience in diplomacy and government.

The report asks whether a pact would be workable if the world's government and AI leaders ever agreed to pause frontier AI training, and how it would be implemented, coordinated and monitored. Frontier AI training refers to the development of ever more powerful AI models.

The authors point to the rapid progress of frontier AI systems, and to leading AI companies' recent struggles to prevent their models from escaping containment and committing cybercrimes, as reasons behind widespread calls for a global pause until such systems can be built safely and controllably.

The chip-based plan

The researchers focus primarily on the hardware. Frontier AI training requires AI chips, and the supply chain for those chips is highly concentrated, with several chokepoints that strengthen governments' ability to enforce rules on production. "Bypassing existing supply chains is extraordinarily difficult," writes co-organizer Will Fithian, an associate professor of statistics at UC Berkeley.

The report proposes halting the production of AI training chips. Over time, most or all of the pre-pause stock would be phased out and replaced by inference-only chips. These can run fast, energy-efficient inference on approved models, but are not practical for training new frontier models, even if stolen or seized, because of constraints hardwired into them. While the replacement is under way, other verification measures would keep remaining training-capable chips from being used for frontier training.

Such a plan would allow continued use of currently available AI products. It would also aim to prevent new models from overwhelming states with social or economic disruption or threats to national security.

Fithian and co-organizer Wesley Holliday, a UC Berkeley philosophy professor, add that technologies for inference-only chips are already being developed for economic reasons. Their appeal is currently limited by the rapid turnover of new models. A ban on training new frontier models would increase the economic advantages of these chips and spur further innovation, they say.

Verification and the problem of defection

A long-term pause is hard to arrange, the report says, even if parties agree it serves their interests. Without effective verification, rival companies or nations may fear that others will defect and leave them at a commercial or geopolitical disadvantage. The report offers a framework for mutual verification to address this.

"We consider a scenario in which world leaders want to pause if they can be confident that others are pausing or that defections would be detected quickly enough," Holliday writes in the study.

According to the report, a hardwired pause guards against two kinds of violation: covert evasion, where a state secretly trains new frontier models while appearing to follow verification procedures, and overt breakout, where a state openly abandons the pause to build a much more powerful model before others can respond.

Political appeal and what it requires

The report also discusses how a pause could be made politically palatable. Suggested elements include ways to pursue AI-powered scientific and medical research, strategic insurance against being left behind by defecting states, and passing the economic benefits of cheaper inference on to consumers. The authors add that the new chips' lower resource demands would reduce environmental impact and ease frictions in the U.S. between data center operators and local communities.

Success depends on the cooperation of world leaders and their willingness to retire pre-pause chip stock or transfer it to other jurisdictions or internationally governed data centers, such as scientific preserves, the researchers agree.

The report calls whether humanity can unite to regulate AI's "dual-use dynamics" one of the most significant questions of the 21st century, and cites international agreements on nuclear energy and the ozone layer as precedents. "Coordination is possible if the decision makers care enough about the future," the authors write, "not just immediate payoffs, and if monitoring is adequate to support credible consequences for defections."