Should governments have kill switches for all AI systems?

A “kill switch” in the context of artificial intelligence refers to a built-in mechanism that allows an AI system to be safely and immediately shut down or deactivated by human operators. This concept is closely associated with safety protocols in advanced or autonomous AI systems, particularly those capable of operating independently or making decisions without direct human oversight. The term originates from computer science and engineering, where emergency stop functions have long been standard in machinery and robotics. In AI, the idea of a kill switch has evolved to address concerns about unpredictability, self-learning behaviors, and the potential for unintended outcomes. It is often discussed in the context of autonomous agents, machine learning systems, and reinforcement learning environments—especially when these systems operate in critical sectors like defense, finance, or infrastructure. Governments and research institutions began seriously discussing AI kill switches in the mid-2010s, as public interest in AI safety grew alongside rapid advances in machine learning. Organizations such as the Future of Life Institute and research labs like DeepMind have contributed to academic work on safe interruptibility—the technical challenge of ensuring an AI system does not resist shutdown or alter its behavior when it detects an override.