Policy
Four Verifiable Constraints for Governing Frontier AI Development
A new framework proposes licensing, weight security, and isolated environments to enable U.S.-China cooperation on AGI safety.
Key takeaways
- Frontier AI development requires industrial-scale compute that is difficult to hide and can trigger licensing through observable thresholds of chip usage and electricity consumption.
- Model weights can be secured through encryption, tamper-evident logging, and declared inventories subject to inspection similar to nuclear material protocols.
- Testing and evaluation of frontier models should occur in isolated replica environments with no external network access to prevent loss of control during capability assessment.
- Internal deployment of models to accelerate successor development poses recursive self-improvement risks requiring continuous monitoring and human oversight.
- The framework proposes a five-level legal structure from bilateral U.S.-China agreement through UN Security Council resolution to create verifiable international restraints.
A Framework for Verifiable AI Restraint
Legal scholars at Just Security have outlined a detailed framework for governing frontier artificial intelligence development through four verifiable constraints, challenging the assumption that U.S.-China cooperation on AGI safety is impossible to verify.
The proposal addresses a core objection to international AI cooperation: that restraint cannot be verified when nations develop advanced AI systems. The authors argue that frontier model development is an industrial-scale process with observable features that can be constrained, reported, and independently audited—at least in the near term.
The Four Pillars of Constraint
The framework centers on four elements of frontier AI development that can be subject to verifiable restraint:
Compute capacity would trigger a licensing regime. Training frontier models currently requires approximately 10²⁶ floating-point operations, demanding tens to hundreds of thousands of specialized chips running continuously for months at costs exceeding hundreds of millions of dollars. Such operations consume electricity on the scale of a small city and are difficult to hide. Any facility approaching specified compute thresholds would be presumed to be developing frontier models and required to obtain a license. Illicit development would be detectable through electricity grid filings, satellite observation, and chip tracking—since suppliers are extremely limited and chips could be designed to reveal their location.
Model weights—the trained parameters that constitute the finished model—would be regulated through two mechanisms. First, a prohibition on publishing or disclosing weights of models above certain capability thresholds. Second, mandatory security standards for weight custody, including encrypted storage with decryption keys released only to licensed chips, tamper-evident logs of all access and transfers, and declared inventories subject to inspection similar to International Atomic Energy Agency protocols for nuclear material.
Model environments would require increasingly isolated and secure conditions across four development phases: training, testing, evaluation, and internal deployment. The highest-risk activities—particularly cyber capability evaluations—would be conducted in closed replica environments with no external network access. Internal deployment, where models are used to accelerate development of successor systems, poses recursive self-improvement risks and would require continuous monitoring and human review of all proposed changes to training processes.
Model capabilities would be assessed through standardized testing, though the authors acknowledge capability alone cannot serve as the initial licensing trigger since it is not externally verifiable without access to the model itself.
Why It Matters
This framework offers a concrete alternative to the prevailing assumption that an AI arms race is inevitable. By identifying specific, verifiable constraints on frontier development, it creates a foundation for mutual restraint between the United States and China—what the authors call a "stag hunt" scenario where cooperation becomes each state's rational response when they can verify the other's compliance. The proposal recognizes that some verification mechanisms will become more difficult over time as compute requirements decrease, making near-term action urgent. Without such frameworks, frontier AI development proceeds in conditions that risk catastrophic loss of control for developers and potentially everyone else.
Implementation Through Five Legal Levels
The authors propose implementing these constraints through a five-level legal structure: a bilateral U.S.-China political agreement; domestic licensing regimes in each state; an independent international standards and inspection body; regulation of the chips required for frontier training; and ideally a UN Security Council resolution extending the regime globally.
The framework acknowledges challenges, including that China would need to abandon its current practice of releasing open-weight frontier models, and that compute-based licensing thresholds must remain flexible as efficiency improves. The authors note verification will become harder over time, reinforcing the case for urgency.
These details were first reported by Just Security in the second part of a two-part essay on frontier AI governance.