A capability planning approach for AI crisis preparedness

Drawing on capability-based planning from defense and homeland security, the researchers propose a methodological framework for government AI crisis preparedness to replace the predict-then-act paradigm. The authors note that AI is difficult to predict, that expert judgments on AI timelines differ by several orders of magnitude, and that official reviews acknowledge that likelihood-based risk assessment fails precisely on this class of risks. The framework has three parts: systematically sampling along declared axes to build a scenario library; evaluating government capabilities against each scenario item by item using coarse-grained threshold-style criteria; and mapping the resulting matrix into decision rules that governments can adopt to set priorities. The authors demonstrate the insights the tool can provide through a pilot targeting four categories of the most severe AI-enabled threats, as a proof of concept that capability planning can be used for AI crisis preparedness.