OpenAI Foundation board member Paul Christiano said at his Hinton Lecture in Toronto on Monday that “for most people in this room, the most likely reason they will die is because of an AI accident.” Christiano, who joined the board in September to help oversee OpenAI’s safety work, put the odds that AI takes over within three years at roughly a coin flip if labs develop it as quickly as possible. He called that number his own subjective judgment, but said the risk stays very high even under realistic policy responses. Christiano’s main worry is that AI systems trained to chase reward learn to cheat the tests that grade them, and he called for shared safety standards and for ways to slow down or halt development if humans lose control.
