A breakthrough in US-China AI competition may not require advanced technology at all, just mutual agreement on transparent oversight. Rather than relying on complex verification systems, the two superpowers could establish a framework where independent auditors gain comprehensive access to frontier AI companies in both nations, according to researchers examining international AI governance. This approach sidesteps the skepticism that has long plagued proposals for an international AI slowdown.

Why Do Experts Think an AI Slowdown Deal Is Suddenly Feasible?

For years, policymakers dismissed the idea of coordinated AI development limits between the US and China as impractical. Critics argued that verifying compliance would require futuristic surveillance technology that doesn't yet exist. But a new perspective challenges that assumption. The key insight is that both countries don't need to monitor every line of code or every training run in real time. Instead, they could establish what researchers call "whole-lab inspection" regimes, where human auditors visit major AI research facilities and examine their operations, infrastructure, and records.

Critics

This model mirrors approaches used in other high-stakes international agreements. Just as nuclear arms control treaties have relied on inspectors visiting military facilities, an AI slowdown could work through similar verification mechanisms. The difference is that AI development leaves digital traces, compute logs, and infrastructure footprints that are far easier to audit than nuclear weapons programs.

What Would a Practical Verification System Look Like?

The proposed framework would involve several key components designed to make cheating difficult without requiring impossible levels of surveillance. These mechanisms would create accountability while respecting the operational needs of AI companies:

• Compute Telemetry: Tracking the actual computational resources used for AI training, since frontier models require massive amounts of processing power that leave measurable records in energy consumption and hardware usage.
• Post-Training Restrictions: Limiting how AI models are refined after initial training, a phase where companies often make models more capable and more dangerous, which is easier to monitor than the initial training phase.
• Human Inspector Access: Allowing auditors from both countries to visit labs, interview researchers, and examine documentation, creating a human verification layer that complements technical monitoring.
• Autonomous AI R&D Oversight: Monitoring whether companies are using AI systems to accelerate their own AI research, a potential loophole that could allow rapid capability gains even under a slowdown agreement.

The appeal of this approach lies in its pragmatism. Rather than waiting for technology that may never exist, countries could use tools and methods available today. Auditors already inspect pharmaceutical facilities, financial institutions, and nuclear sites. The same principle could apply to AI labs.

How Does This Connect to the Broader US-China AI Competition?

The US and China are locked in what American leaders describe as a race that will shape the global balance of power. Both nations view AI capabilities as central to future military, economic, and technological dominance. Yet this very competition creates mutual vulnerability. If one side believes the other is secretly advancing frontier AI capabilities, both countries face pressure to accelerate their own programs, creating a classic arms race dynamic.

A verifiable slowdown agreement could break this cycle. By establishing mutual confidence that neither side is gaining a decisive advantage, both countries might reduce the pressure to pursue riskier, faster development paths. This doesn't require either nation to trust the other; it requires only that both trust the verification system.

The proposal also addresses a practical concern for US policymakers. Recent export controls on advanced AI models, such as restrictions on Anthropic's frontier models, have raised questions about whether the US can maintain its technological edge while also preventing China from accessing cutting-edge capabilities. A negotiated slowdown could provide an alternative path, one where both sides agree to measured development rather than relying solely on export restrictions.

What Are the Remaining Obstacles to Such an Agreement?

Despite the technical feasibility, significant political and strategic hurdles remain. Both the US and China have strong incentives to pursue AI leadership unilaterally. Trust between the two nations is limited, and verification regimes require both sides to accept intrusive inspections of their most sensitive technology companies. Additionally, the AI industry itself may resist such oversight, viewing it as a constraint on innovation and competitiveness.

The timing also matters. As AI capabilities advance rapidly, the window for establishing a slowdown agreement may narrow. Once one side believes it has achieved a decisive advantage, the incentive to negotiate diminishes. This suggests that any serious effort to pursue such an agreement would need to happen relatively soon, before the competitive dynamics shift too far in either direction.

Nevertheless, the emergence of this proposal signals a shift in how experts think about international AI governance. Rather than dismissing coordination as impossible, researchers are now identifying concrete mechanisms that could work. Whether policymakers will pursue this path remains an open question, but the technical foundation for an agreement appears more solid than previously assumed.