Experts explain why an emergency button is not enough to shut down AI

An AI can be stopped when it is in a controlled environment, but the task becomes more complex as the system gains autonomy and access to other programs, files and machines. Nicolas Papernot, a professor at the University of Toronto, told AFP that an operator can cut off access to the computing power it needs; if an agent spreads to other devices, its copies would have to be stopped one by one.

Thierry Poibeau of France’s National Centre for Scientific Research (CNRS) says there is no button capable of shutting down AI as a whole, although specific systems can be stopped. In a LinkedIn article, Hussein Abbass of the University of New South Wales (UNSW) described increasingly difficult levels as a system moves beyond its original environment. Jean-Gabriel Ganascia said the more concrete risk may be growing dependence on these systems, whose interruption could also affect legitimate services.