Can AGI Rivals Hold Back? Verification, Domestic Oversight, and the Conditions for Mutual Restraint

RAND researchers use a two-stage game-theoretic model to identify when rival states can sustain mutual restraint in AGI development without observing each other's compliance. The topline: 'Without domestic AI oversight, rivals can't credibly commit to restraint—even when both sides want to.' Key findings: fear of catastrophic risk only partially deters cheating (purely positional defections require detection); verification loses force as AGI arrival appears imminent; and mutual doubt can trap rivals in an opaque world. Recommendations treat domestic AI governance — compute registries, licensing, facility audits, incident reporting — as security architecture rather than mere regulation, and urge investment in confidential-computing and trusted-hardware verification plus instruments tying post-AGI outcomes to pre-AGI conduct (published Sep 30, 2026).

This is the most rigorous recent articulation of why compute registries and domestic oversight matter for international AI restraint — directly relevant to any enterprise or investor exposed to compute-governance policy.

Fold the verification/oversight logic into your AI-risk and compute-asset planning, as compute registries look set to become binding security architecture.