When nations view AI as a zero-sum game, they often treat safety protocols as friction that slows down their rivals. In this mindset, the first country to deploy a powerful model wins, even if that model is unstable. To counter this, treaties must move beyond vague promises of cooperation and focus on technical verification and shared risks.
One practical path involves creating international monitoring bodies, similar to how the IAEA manages nuclear materials. Instead of asking countries to 'be safe,' these frameworks should mandate inspections of massive compute clusters or require hardware-level telemetry that flags unauthorized training runs. If a nation knows its neighbors are being monitored by neutral technical observers, the incentive to cut corners diminishes. We need technical proofs, not just diplomatic platitudes.
Furthermore, global governance can create
