One master AI kill switch is probably too risky. We need multiple layered kill switches combined with software solutions. One idea I think is important builds on the philosopher Yuval Noah Harari’s warnings about AI deception. He argues for rules that keep AI from masquerading as human. I take that a step further: we should build human laws and computer rules that do not allow AI to deliberately lie. Other ideas include limiting permissions, monitoring behavior, and requiring human approval for important actions. This tidbit focuses on kill switches.
Data sitting on a drive cannot physically do anything. The bigger risk comes when AI can reach into the material world: controlling things like computers, financial accounts, and autonomous robots. The more capable AI becomes, the more important it is to control the bridges between information and physical action.
The Internet itself is one of those boundaries. Militaries already use separate networks, but they are not always fully air-gapped (separated physically from the public internet). So, more air gapping is part of this, not just better segmentation. Future safeguards should allow operators to disconnect a data center, cloud region, or perhaps even large portions of the public Internet. The tension, of course, is free speech versus public safety.
Then there are physical choke points. AI still needs electricity. We need the ability to disconnect data centers, and other network facilities, from power, and this should include an ability to disconnect backup generators. It’s clear, no single safeguard is enough. If one containment layer fails, humans need to have another door they can close. And to protect humanity from AI disasters, I wonder if ultimately, we should start working toward the ability to temporarily shut down electrical power worldwide — at least as completely as physically possible.