Frontier AI development is concentrating in a small number of labs running clusters that cost billions to build and that outside institutions cannot currently observe. This unit makes the case that governance claims only matter when someone can check whether they are true.
Start with compute as a governable input, ground the argument in the current AI risk consensus, then use nuclear arms-control verification as a historical anchor for what mature verification regimes actually require.
The canonical argument that compute is more governable than data, talent, or algorithmic insight. Read for the strategic frame: AI governance worth building depends on whether claims about compute can be verified.
→ arXiv:2402.08797The closest the field has to a consensus risk statement written for governments. Read the risk sections to understand what verification is ultimately trying to make governable.
Available at gov.uk/government/publications/international-ai-safety-report-2025.If alignment were solved, the case for external verification would be weaker. Read for the open problems that make assurance through behaviour alone unreliable.
→ arXiv:2404.09932A concrete account of treaty verification in another high-stakes domain. Read for institutional architecture, on-site inspection, seals, failure modes, and which lessons transfer to AI.
→ arXiv:2304.04123