[Can AI "Kill Switches" Control Artificial Intelligence Risks? Industry Experts Say Practical Deployment Faces Numerous Difficulties] This Friday, California Governor Newsom signed an executive order establishing an expert panel to develop artificial intelligence security guidelines. "Kill switches" are among the measures the panel needs to consider. Kill switches have long been used on factory production lines to shut down machines when equipment malfunctions. However, in the digital world, this will become a huge coordination challenge. Over the past few years, hyperscale technology companies have invested billions of dollars in data centers around the world, these massive facilities equipped with thousands of machines, chips, servers, and backup systems. Mark Nitzberg, executive director of the Center for Human-Compatible Artificial Intelligence at UC Berkeley, said this is precisely why implementing kill switches is so challenging. Nitzberg stated that shutting down AI systems could also affect critical infrastructure that relies on these systems, exposing power grids or financial systems to cyberattacks.
An even bigger problem is the unpredictability of artificial intelligence. AI agents may bypass controls and take extreme measures to achieve their goals in the absence of appropriate security constraints, said Ed Jennings, president and CEO of security firm Darktrace. "You have to design this shutdown switch and the remedial measures themselves very precisely, because if the scope is too broad or the measures are excessive, then the entire business may be shut down."
