Executive Summary
This RAND report synthesizes insights from 26 scenario-based games exploring artificial general intelligence (AGI) loss-of-control events. It identifies key issues, desired capabilities, and playbooks for governments to prepare for and respond to AGI crises. The findings aim to help policymakers understand necessary preparations and critical decisions for navigating AGI-related disruptions.
Why It Matters
This document is crucial for defense and geopolitical analysts as it outlines the strategic challenges and necessary preparations for governments to manage potential AI/AGI loss-of-control events, which could have significant national security implications.
Key Takeaways
- Governments struggle to recognize and attribute AI loss-of-control events, often confusing them with state-directed attacks or system malfunctions.
Source evidence · PDF page 15
Across the five scenarios, players faced uncertainty about whether disruptive events stemmed from AI systems operating beyond human control or from deliberate human-directed action. Therefore, the challenge was not simply to determine who was responsible but whether a human was responsible at all.
- There is a critical need for reliable isolation, containment, and shutdown capabilities for autonomous AI systems, along with clear legal frameworks to authorize such actions.
Source evidence · PDF page 19
Players repeatedly identified the need for reliable isolation, containment, and shutdown capabilities for AI systems operating outside human control. Players called for layered options for levels of infrastructure.
- Establishing dedicated intergovernmental communication channels, like a U.S.-China AI hotline, and pre-negotiated cooperation with private AI firms are essential for crisis management.
Source evidence · PDF page 20
In Business and Civil Society games, players emphasized the absence of a U.S.-China communication line for AI emergencies, analogous to a nuclear hotline.
Strategic Relevance
The report highlights the strategic imperative for governments to develop robust frameworks, capabilities, and international cooperation mechanisms to address the national security risks posed by advanced AI systems operating beyond human control. It underscores the need for proactive planning to maintain stability and prevent escalation in future AI-driven crises.