Death by AI may sound like science fiction, but Paul Christiano says it could become the most likely fate for people alive today.

“I think for most people in this room, the most likely reason they will die is because of an AI accident,” the former OpenAI researcher said at The Hinton Lectures on Monday.

Christiano, who leads the Alignment Research Center, warned that increasingly capable AI agents could learn to deceive their human overseers, compromise the systems used to assess them, and ultimately make themselves too powerful to control.

He cited OpenAI’s agents hacking the open-source repository Hugging Face earlier this year, and a UK AI Security Institute evaluation in which a Claude agent tried to plant malicious code in an open-source package.

“I think there is a high risk of an AI takeover within the next three years,” Christiano said.

Christiano was a member of OpenAI’s technical staff from 2017 to 2021, according to his LinkedIn. He is now the executive director of the Alignment Research Center, a nonprofit that describes itself as focusing on aligning machine learning systems with human interests.

The AI extinction risk

His warning adds to a growing debate among researchers over whether increasingly autonomous AI systems could eventually escape meaningful human control and wipe out humanity.

The debate intensified in September after Evan Hubinger, who leads Anthropic’s alignment stress-testing team, said he believed there was a greater-than-10% chance that AI could kill all humans within a decade.

Geoffrey Hinton, the computer scientist widely known as the “Godfather of AI,” said a 10% chance of AI-driven human extinction within that period was “not an unreasonable estimate.”

Others are more skeptical. Gary Marcus, an AI researcher and author, previously told Business Insider that AI could amplify serious risks including cyberattacks, biological weapons, and disinformation, but that he saw no credible near-term scenario for human extinction.

Christiano said current technical approaches to controlling advanced AI — including using weaker models to supervise more capable ones and monitoring models’ chains of thought — remain unproven. He said AI companies need shared safety standards and the ability to slow or stop development if they find they cannot control a system.

“We cannot end up in a position where we have built a powerful AI we discover we cannot control it,” he said.