Source-linked AI summary
Governing dual-use technologies: Case studies of international security agreements and lessons for AI governance
Akash R. Wasil, Peter Barnett, Michael Gerovitch, Roman Hauksson, Tom Reed, Jack William Miller
TL;DR
Advanced AI presents global security risks, and many such risks are inherently international. The paper draws lessons from international security-agreement case studies, highlighting design considerations for AI governance while identifying issues that remain for future work.
Problem
Advanced AI presents major security challenges, and many AI risks are inherently international in nature.
Source: Introduction · p. 4Method
The paper draws lessons from international security-agreement case studies, examining each agreement's purpose, core powers, governance, and non-compliance.
Source: IAEA (International Atomic Energy… · p. 3, Introduction · p. 4Results
The case studies identify verification, balanced representation, adaptability, transparency trade-offs, and benefit-sharing as lessons for international AI governance.
Source: IAEA (International Atomic Energy… · p. 3Takeaways & Limitations
Future AI governance should consider robust verification, representation of global interests and leading AI powers, and incentives such as sharing safe and responsible AI applications.
Source: IAEA (International Atomic Energy… · p. 3Further work is needed on AI-specific verification, governance decisions, technical expertise, and strategies to prevent or respond to non-compliance.
Source: Future work · p. 3
Abstract
from arXiv · showhide
International AI governance agreements and institutions may play an important role in reducing global security risks from advanced AI. To inform the design of such agreements and institutions, we conducted case studies of historical and contemporary international security agreements. We focused specifically on those arrangements around dual-use technologies, examining agreements in nuclear security, chemical weapons, biosecurity, and export controls. For each agreement, we examined four key areas: (a) purpose, (b) core powers, (c) governance structure, and (d) instances of non-compliance. From these case studies, we extracted lessons for the design of international AI agreements and governance institutions. We discuss the importance of robust verification methods, strategies for balancing power between nations, mechanisms for adapting to rapid technological change, approaches to managing trade-offs between transparency and security, incentives for participation, and effective enforcement mechanisms.
Executive Summary
The case studies compare how international security agreements distribute powers, govern participation, and handle non-compliance. Their lessons for AI governance emphasize verification, balanced representation, adaptability, transparency, and incentives to participate.
IAEA (International Atomic Energy Agency)
- The IAEA promotes peaceful nuclear energy, prevents proliferation, and verifies treaty compliance through facility inspections and reports to the UN Security Council; its 35-member Board includes 13 capability-preselected members and 22 elected for two years.The effectiveness of the 2015 Iran nuclear deal was questioned and undermined when the US withdrew.Source: IAEA (International Atomic Energy… · p. 2
START Treaties
- START treaties use inspections, satellite monitoring, and data exchange to reduce US and Russian arsenals, but Russia stopped allowing inspections in 2023.The bilateral Consultative Commission makes decisions by consensus.Source: IAEA (International Atomic Energy… · p. 2
The Organisation for the Prohibition of Chemical Weapons (OPCW)
- The OPCW inspects chemical-weapons elimination and can suspend rights and privileges after non-compliance, as it did after Syria used chemical weapons and failed to comply with inspections.Its 41-member Executive Council is elected by the Conference of the States Parties.Source: IAEA (International Atomic Energy… · p. 2
Wassenaar Arrangement
- The Wassenaar Arrangement coordinates export-control policies through consensus, which can hinder updates, as illustrated by Russia's unilateral blocking of list revisions.Annual meetings and working groups support its policy and technical discussions.Source: IAEA (International Atomic Energy… · p. 2
Biological Weapons Convention (BWC)
- The BWC relies on state commitments and voluntary information-sharing without formal inspections, and failed to detect the Soviet Union's secret biological-weapons program.Its decentralized Review Conferences discuss scientific developments, while states address potential violations through consultation and cooperation.Source: IAEA (International Atomic Energy… · p. 2
Key Findings and Lessons Learned
- The lessons for AI agreements include robust verification, balanced representation, adaptability to technological change, transparency trade-offs, and benefit-sharing to encourage participation.The proposed AI applications include technical expertise to track advances and sharing safe, responsible AI uses.Source: IAEA (International Atomic Energy… · p. 3
Future work
The paper identifies unresolved questions that require further work to make international AI governance mechanisms more effective and robust.
Future work
- Future work should develop robust, adaptable methods to verify AI development and deployment.The paper also calls for identifying the key decisions an international institution should make and how those decisions would be made.Source: Future work · p. 3
- Future work should design governance structures that balance global representation with the interests of leading AI nations and incorporate technical expertise to keep pace with AI advances..Source: Future work · p. 3
- Further research should examine ways to prevent non-compliance with AI agreements or respond to confirmed violations..Source: Future work · p. 3
Introduction
The paper situates international AI governance in the security risks of advanced AI and the potential role of international cooperation. It reviews security agreements and institutions to derive lessons for future AI governance.
Introduction
- Advanced AI is described as posing global security challenges, with the Bletchley Declaration recognizing that many AI risks are inherently international and best addressed through cooperation..Source: Introduction · p. 4
- Existing coordination includes national AI Safety Institutes and calls for an inclusive international network and coordinated global action on AI safety research and governance.The passage cites UK Government 2024, NIST 2024, and IDAIS 2023.Source: Introduction · p. 4
- Prior proposals include shared safety-focused AI development and institutions that could set standards, monitor compliance, detect unauthorized development, or certify national regulators.The passage cites Maas and Villalobos 2023; Cass-Beggs et al. 2024; Hausenloy, Miotti, and Dennis 2023; Ho et al. 2023; and Trager et al. 2023.Source: Introduction · p. 4
- The paper reviews international security agreements focused on dual-use technologies to identify lessons for agreements or institutions intended to reduce global risks from advanced AI.Its case studies examine each agreement's purpose, core powers, governance, and instances of non-compliance.Source: Introduction · p. 4
2. Lessons learned.
This section states that the paper summarizes lessons from international security agreements for thinking about international agreements concerning advanced AI development or deployment.
2. Lessons learned.
- The paper summarizes lessons from international security agreements for considering agreements related to advanced AI development or deployment..Source: §2 · p. 4
Case studies of international agreements
This section examines international agreement case studies, including their monitoring and governance arrangements and instances of non-compliance, to inform AI agreement design. It focuses on verification, balancing national power, managing transparency and security trade-offs, adapting to technological change, incentivizing participation, and addressing enforcement challenges.
IAEA
- The IAEA combines inspections with escalation to the UN Security Council, while its 35-member Board includes seats reserved for advanced atomic-energy nations.The agency’s stated purpose is peaceful nuclear use and limiting military use; its Director General oversees strategy, policies, staff, and inspector recruitment.Source: IAEA · pp. 4–5
- IAEA inspections exposed Iran’s Amad Plan violations, but the JCPOA’s later U.S. withdrawal and reimposed sanctions were followed by reduced Iranian compliance.The 2015 agreement required Iran to limit nuclear-program investment and permit extensive IAEA monitoring in exchange for sanctions relief.Source: IAEA · p. 5
START Treaties
- Russia’s suspension of New START participation and halt to inspections left compliance difficult to verify, with the treaty set to expire in 2026 absent further negotiations.The treaty includes inspections and data exchanges; both countries stated they would continue adhering to its numerical limits despite the suspension.Source: START Treaties · pp. 5–6
The Organisation for the Prohibition of Chemical Weapons (OPCW)
- The OPCW suspended Syria’s rights after chemical-weapons non-compliance, yet limited enforcement and Russian diplomatic protection left accountability unresolved.The United States said Russian forces used chemical weapons in Ukraine, but the OPCW had not opened an official inspection because evidence was deemed insufficiently substantiated.Source: The Organisation for the Prohibition… · p. 7
Wassenaar Arrangement
- Wassenaar’s consensus-based, voluntary export-control regime lacks formal inspection, enforcement, and expulsion powers, allowing Russia to impede control-list updates.Other nations consequently implemented ad hoc export controls outside the arrangement’s framework.Source: Wassenaar Arrangement · pp. 7–8
Biological Weapons Convention
- The BWC’s lack of robust verification allowed the Soviet Union’s covert Biopreparat program to remain undetected for years.The Convention relies on national implementation and annual Confidence-Building Measures reports, but has no formal compliance verification mechanism.Source: Biological Weapons Convention · pp. 8–9
- Challenge inspections, on-site inspections, and continuous monitoring supported agreement effectiveness, while the BWC’s lack of formal verification enabled non-compliance and weakened strengthening efforts.For AI agreements, these methods could help detect non-compliance.Source: Biological Weapons Convention · p. 9
- Governance can balance geopolitical influence with regional representation, but consensus risks gridlock, bilateral agreements may break down, and key nations may withdraw from commitments.AI agreements could weigh bilateral provisions drawing on START against broader representation with extra decision-making power for AI-leading nations, including the United States and China.Source: Biological Weapons Convention · p. 9
- Agreements must balance transparency with state and commercial interests, as proprietary-information concerns blocked a BWC verification protocol and controlled-item disputes challenged Wassenaar.AI verification could involve inspections, code or data-center access, or automated hardware alerts, while protecting sensitive materials such as dangerous model weights.Source: Biological Weapons Convention · pp. 9–10
- AI governance institutions must track rapid technical change and use technical expertise to interpret evaluations, assess safety cases, anticipate advances, and identify ways to detect non-compliance.Technical staff may be more attracted to institutions with constructive, innovative aims than to a solely policing role.Source: Biological Weapons Convention · p. 10
- Benefit-sharing can incentivize states to join or remain in agreements, and AI governance could promote beneficial research and development while mitigating risks.Examples include OPCW support for peaceful chemistry and IAEA promotion of peaceful nuclear energy.Source: Biological Weapons Convention · p. 10
- Because international bodies may lack direct enforcement powers, AI governance must decide what actions an institution can take itself and how much to rely on the UN Security Council or member nations.Reliance on external enforcement can lead to political deadlocks, as in the case of Syria’s chemical weapons use.Source: Biological Weapons Convention · p. 10