International Agreements to Limit Frontier AI: Objectives and Exit

Technical Summary: International Agreements to Limit Frontier AI: Objectives and Exit

Problem Statement
The paper addresses the critical gap in proposals for international agreements to limit frontier AI development. While existing literature (e.g., Scher et al., 2025; Miotti & Wasil, 2023) focuses heavily on the mechanisms for imposing restrictions (such as computational power limits), it largely neglects the conditions under which these restrictions should be relaxed or the agreement terminated. The authors argue that the lack of a clear, agreed-upon "exit strategy" or relaxation mechanism creates uncertainty for signatories, potentially acting as a barrier to forming such agreements. The core problem is determining the specific criteria that should trigger the transition from a restricted development phase to a relaxed or terminated phase, balancing the need to mitigate existential risks against the potential economic and societal benefits of AI.

Methodology
The authors employ a comparative policy analysis and conceptual framework approach:

1. Survey of Existing Treaties: The paper analyzes termination and withdrawal clauses in eight major international agreements, including the Strategic Arms Reduction Treaty (START), the Non-Proliferation Treaty (NPT), the Biological Weapons Convention, and the Paris Agreement.
2. Definition of Desiderata: The authors systematize the selection of relaxation conditions by proposing four key criteria (desiderata):
   • Accuracy: The condition must accurately reflect whether AI risks have been sufficiently mitigated.
   • Verifiability: Fulfillment of the condition must be unambiguous and provable.
   • Preservation of Sovereignty: Signatories must retain a plausible path to exit if their risk assessments change.
   • Precedent: Conditions should ideally rely on mechanisms with a history of success in international law.
3. Scenario Modeling: The analysis assumes a specific context: an agreement signed by NATO member states and China, imposing an upper limit on computational power for training AI. The authors assume a "pessimistic" scenario where signatories lack the technical infrastructure to unilaterally halt each other's development, relying instead on reputational costs and mutual trust.
4. Evaluation of Candidates: Various types of relaxation conditions are evaluated against the established desiderata.

Key Contributions and Proposed Conditions
The paper categorizes and evaluates several potential types of relaxation conditions:

• Absence of a Relaxation Condition: A permanent ban. The authors note this has low precedent for termination and low sovereignty preservation, though it offers high verifiability.
• Fixed Time Limit: A set duration after which restrictions are reviewed. This scores high on verifiability and precedent but low on accuracy, as it does not account for the actual state of AI safety research.
• Individual Withdrawal / Extraordinary Events: Allowing states to withdraw if "extraordinary events" occur (similar to the NPT and ABM treaties). This preserves sovereignty but suffers from low verifiability and potential for dispute over legitimacy.
• Progress in Preparedness: Releasing restrictions based on advancements in AI safety (e.g., alignment, interpretability) or societal resilience. This scores high on accuracy but faces challenges in verifiability and defining specific milestones (e.g., "provable guarantees" or "expert consensus").
• Absence of AI-Related Incidents: Using the lack of harmful incidents as a signal for relaxation. The authors note this is vulnerable to manipulation and may not guarantee safety under unrestricted conditions.
• Creation of a Global AI Development Project: Requiring the formation of a unified, globally monitored development entity. This addresses coordination failures but requires unprecedented international cooperation and may be unlikely to be satisfied.
• Deference to a Later Specification: A two-stage agreement where the first stage establishes restrictions, and a newly formed organization defines the specific relaxation conditions for the second stage.

Results and Recommendations
Based on the trade-offs between accuracy, verifiability, sovereignty, and precedent, the authors propose a specific hybrid recommendation for an example scenario:

1. Fixed Initial Period: The agreement should operate for a fixed period of 5 years.
2. Establishment of a New Organization: During this period, a new international organization should be established.
3. Specification of Conditions: This organization is tasked with defining the specific relaxation conditions that address preparedness for rogue AI, catastrophic misuse, and societal disruption.
4. Decision Mechanism: After 5 years, signatories must agree on these conditions. If they cannot agree, they may extend the agreement by 2 years or terminate it.
5. Withdrawal Clause: A provision for individual withdrawal based on extraordinary events (e.g., a non-signatory state acquiring frontier capabilities) with a 4-month notice period.

The authors argue that this approach balances the need for a definitive safety assessment (via the new organization) with the necessity of preserving state sovereignty and allowing for future consensus.

Significance
The paper claims its primary significance lies in illustrating the considerations required to make an international AI limitation agreement viable. By shifting focus from the restrictions themselves to the "exit" mechanism, the authors aim to provide a framework that could reduce the political friction preventing such treaties from being signed. The work suggests that without a credible and agreed-upon path to relaxation, signatories may be unwilling to commit to restrictions that could forego significant economic benefits. The authors frame this not as a definitive solution, but as a necessary step in the technical governance of AI to address what they identify as the "single largest risk" facing humanity today.