The Case for AI Loss of Control Response Planning and an Outline to Get Started

As artificial intelligence (AI) systems become increasingly capable and diffused throughout the economy and society, the possibility of loss of control (LOC) incidents in which AI systems operate outside human control may require new national-level responses. Detecting and responding to a significant LOC incident requires not only a technical understanding of AI models but broader whole-of-society planning and coordination with stakeholders in government, companies in the AI tech stack, critical service providers, and end users. However, the existing response environment is underdeveloped, with limited agreement on risks and few coordination mechanisms.

In this paper, the authors explain the case for a national plan for responding to significant LOC incidents and propose criteria and an outline for a plan, with the goal of designing the practical processes to contain and recover from LOC incidents that reach levels of national significance. First, they explain the case for national LOC response planning. Second, they propose criteria and a structure for national LOC response planning. Third, they provide an outline for a U.S. government plan for LOC incidents, as well as a draft AI LOC severity schema for response planning. The intended audience of this paper includes U.S. government officials, members of private firms, critical infrastructure owners and operators, and leaders of research organizations throughout the AI ecosystem who are important for emergency response planning.