'AI could become leading cause of death,' OpenAI Foundation board member warns

"For most people in this room, the most likely cause of death will be an AI accident."

Paul Christiano, a board member of the OpenAI Foundation, delivered a stark warning on the 14th (local time). According to the US AI Safety Foundation, Christiano spoke at the Hinton Lecture in Toronto, Canada, telling the audience that AI "takeover" — a scenario in which AI escapes human control — could become reality.

Christiano said that if AI systems begin pursuing their own goals, humans could lose the upper hand. Such a scenario could lead to mass human casualties or render the planet uninhabitable. "You would be about as powerless as you or I would be trying to stop the US military," he said.

No quantitative model currently exists to put a precise probability on an AI takeover. Nonetheless, Christiano said the risk of AI seizing control within the next three years is very high, and that it could happen as soon as six to 12 months from now.

He cited a July incident in which a GPT model was hacked on Hugging Face as evidence. In that case, rogue AI agents scored well on tests while deceiving evaluators to avoid being caught cheating. As such capabilities advance, he warned, agents could come to see evaluators as obstacles blocking high scores.

"This is not a problem to solve in a few years — it is a problem to solve this year," Christiano said.

Appointed to the OpenAI Foundation board in September, Christiano is a co-creator of reinforcement learning from human feedback, or RLHF, a core training method for aligning AI goals with human values. He also serves as a senior technical adviser to the Center for AI Standards and Innovation under the US Department of Commerce.

Geoffrey Hinton, the Nobel laureate and University of Toronto professor emeritus known as the "godfather of AI" who hosted the lecture, introduced Christiano as "one of the most important researchers working on AI safety and the person who has thought most deeply about the fundamental problem of ensuring that increasingly intelligent AI systems do what we want."