The fundamental safety risk of AI lies in humanity’s pursuit of machine intelligence superior to its own, without knowing how to maintain control over it. That’s the core argument that leading AI researcher Stuart Russell (UC Berkeley) makes in his DLD26 conversation with Kenneth Cukier (The Economist).
Modern models already show troubling behaviors, Russell notes. “If you put the AI system at risk, in the sense that it has reason to believe it’s going to be switched off, it will do pretty much anything to prevent that.” In test scenarios, this can go as far as allowing a human being to die or launching a nuclear attack, Russell says.
“This seems like a big flashing red light”, the computer scientist warns. “We do not know how to control these systems.”
Governments should regulate the technology more stringently, Russell argues, but “they’re almost paralyzed by the rate of change and what’s coming at them. And at the moment, they are failing to protect the human race.”
Looking back at the history of AI development, Russell says the field’s foundational mistake was treating Artificial Intelligence as an agent pursuing fixed objectives.
“The answer is to say, ‘Actually, the objectives are in the humans,’”, Russell argues. “The AI system should not pursue its [own] objectives, period. It should pursue the interests of human beings.”
Watch the video to explore this session in detail.
