Research · Georgetown

CSET maps where AI kill switches could operate and what they trade off

A new CSET analysis explains how emergency interventions could act across the AI stack, from access controls and networks to chips, data centers, and power.

Georgetown’s Center for Security and Emerging Technology examines how an AI “kill switch” could work at different layers of the technical stack.

Jessica Ji, Jacob Feldgoise, Kyle Miller, Andrew Lohn, Pablo Chavez, and Julie George distinguish narrow interventions, such as restricting an API or blocking an IP address, from broader actions, such as disconnecting chips or data-center power. They argue that the right intervention depends on the incident, the technical owner of the control, and the collateral disruption a shutdown could create.

Why Hoyas.ai is following it

The analysis moves a prominent policy phrase into operational terms. It helps governance, cybersecurity, and public-sector audiences ask who can intervene, how precisely, and how the outcome would be evaluated.

Evidence boundary

This is CSET analysis, not an official Georgetown University policy position. The article describes possible intervention mechanisms and policy considerations; it does not establish that Georgetown, a government agency, or an AI provider has adopted a specific kill-switch system.

Live event calendar

Keep upcoming events on your calendar.

Subscribe once to receive new events and corrections automatically. Your calendar app controls how quickly updates appear.

Download a one-time snapshot if your calendar app does not support subscriptions.