To identify the strategic requirements for mutual restraint, the authors developed a game-theoretic model to examine the competing concerns over competition and cooperation between rival states racing toward artificial general intelligence (AGI). The results show both the conditions under which restraint is feasible and sustainable and how these conditions shift as states come to believe that they are close to attaining AGI.

Can AGI Rivals Hold Back?

Verification, Domestic Oversight, and the Conditions for Mutual Restraint

Tobias Sytsma, Alvin Moon

ResearchPublished Sep 30, 2026

This research was independently initiated and made possible by RAND’s own resources, including income from operations, unrestricted gifts from RAND supporters, and distributions from its long-term investment fund.

Topline

Without domestic AI oversight, rivals can’t credibly commit to restraint—even when both sides want to.

The race to develop artificial general intelligence (AGI) poses a familiar strategic dilemma: Each rival might prefer that everyone hold back, but none can be sure the others will. Any restraint arrangement depends on verification—and verification depends on whether states have built the domestic institutions that make their own artificial intelligence (AI) ecosystems legible.

RAND researchers developed a game-theoretic model to identify the conditions under which mutual restraint is feasible and self-enforcing. They examine how domestic oversight, the detectability of cheating, proximity to AGI arrival, and fears of catastrophe interact to determine whether restraint between rivals can hold.

Key Takeaways

Domestic AI oversight can enable international restraint.

  • A state that hasn’t made its AI industry transparent—even to itself—can’t prove it’s complying with a restraint arrangement.

Fear of catastrophe deters some cheating, but not all.

  • A state that expects to share in the catastrophic risk it creates by defecting is partly self-deterred. But deterring purely positional defections—those aimed at reaching AGI first—requires detection.

Verification loses force as AGI arrival appears imminent.

  • As rivals come to believe AGI is near, the remaining time to punish defectors shrinks. Only instruments tied to post-AGI outcomes may still deter cheating.

Mutual doubt can trap rivals in an opaque world.

  • Even if preferable, restraint may still fail when one rival doubts that the other will make its actions verifiable. Building the means for verification enables states to choose mutual restraint.

Recommendations

To preserve the option of mutual restraint, states should:

  • Treat domestic AI governance—compute registries, licensing, facility audits, and incident reporting—as security architecture, not just regulation, because it creates the evidentiary foundation that restraint arrangements require.
  • Invest in verification technologies that do not expose sensitive information, such as confidential-computing and trusted-hardware methods, to lower the security cost of demonstrating compliance to a rival.
  • Develop instruments that tie post-AGI outcomes to pre-AGI conduct before they are needed, while detection-backed punishment can still reinforce them.
  • Consider reducing a rival’s cost of building oversight capacity through technical assistance or shared monitoring platforms: The feasibility of mutual restraint is determined by whichever rival finds transparency more costly.

Methodology

RAND researchers developed a two-stage game-theoretic model to analyze the conditions under which rival states can sustain mutual restraint in AGI development without being able to directly observe the other’s compliance. In Stage 1, each state chooses how much oversight to impose on its own AI industry beyond what domestic policy requires. In Stage 2, states interact repeatedly until AGI arrives, choosing whether to comply, race openly, or covertly defect in each period. The analysis draws on repeated game theory, arms control literature, and prior RAND research on the strategic dynamics of AGI competition.

Topics

Document Details

Citation

Chicago Manual of Style

Sytsma, Tobias and Alvin Moon, Can AGI Rivals Hold Back? Verification, Domestic Oversight, and the Conditions for Mutual Restraint. Santa Monica, CA: RAND Corporation, 2026. https://www.rand.org/pubs/research_reports/RRA5163-1.html.
BibTeX RIS

Research conducted by

This work was independently initiated and conducted by the Center for the Geopolitics of Artificial General Intelligence within RAND Global and Emerging Risks using income from operations and gifts from RAND supporters.

This publication is part of the RAND research report series. Research reports present research findings and objective analysis that address the challenges facing the public and private sectors. All RAND research reports undergo rigorous peer review to ensure high standards for research quality and objectivity.

This document and trademark(s) contained herein are protected by law. This representation of RAND intellectual property is provided for noncommercial use only. Unauthorized posting of this publication online is prohibited; linking directly to this product page is encouraged. Permission is required from RAND to reproduce, or reuse in another form, any of its research documents for commercial purposes. For information on reprint and reuse permissions, please visit www.rand.org/pubs/permissions.

RAND is a nonprofit institution that helps improve policy and decisionmaking through research and analysis. RAND's publications do not necessarily reflect the opinions of its research clients and sponsors.