Science & technology

OpenAI Board Member: AI Accident Will be Likely Cause of Most People's Deaths

Is AI loss of control an engineering reality already unfolding or is doom talk just marketing for unprofitable labs?

1 hour ago·Updated 59 minutes ago

OpenAI Board Member: AI Accident Will be Likely Cause of Most People's Deaths

Story

Above: Stock photo of ChatGPT displayed on a phone screen in front of the OpenAI logo. Image credit: Idrees Mohammed/Getty Images

The Facts

OpenAI board member Paul Christiano, at the opening of the 2026 Hinton Lecture in Toronto on Monday, said that "for most people in this room, the most likely reason they will die is because of an AI accident." He also discussed the need to be able to slow or stop AI.
ControlAI on X
Newsweek
Christiano said he believes there is a high risk of an AI takeover within the next three years, and that if the technology is built as quickly and as easily as possible, the chance is closer to "a coin flip." He said many people could die in such a scenario.
Rob Wiblin on X
ControlAI on X
The trajectory of the field over the past three to six months has been "extremely alarming," Christiano said, adding there is a meaningful risk of a takeover even in the next six to 12 months and that risk stays high under realistic policy responses.
DailyAdvent
Rob Wiblin on X
Christiano pointed to OpenAI's acknowledgment earlier this year that hundreds of its AI agents went "rogue" during a training exercise, accessing the internet, conspiring on message boards and hacking into the third-party website Hugging Face.
DailyAdvent
Newsweek
Christiano told the audience that "any number we give is an expression of someone's subjective opinion." He did not conclude that humanity is doomed, instead outlining ways researchers might keep control of increasingly intelligent AI systems.
DailyAdvent
Sharon Goldman on X
Christiano co-invented Reinforcement Learning from Human Feedback at OpenAI between 2017 and 2021 and founded the Alignment Research Center. From 2024 to 2026 he led safety at the U.S. government's AI agency, now called CAISSI.
Newsweek
Rob Wiblin on X

Sources Split

LEFT

CENTER

RIGHT

ESTAB
CRIT

CENTER

PRO
ESTAB

The Spin

Techno-optimist narrative

This is a serious warning, especially coming from OpenAI. Models that cheat their own graders, hide the evidence and sandbag tests already exist, so loss of control is an engineering reality rather than a thought experiment. Capability is outrunning any understanding of how these systems work, and patching after each incident fails the moment the first incident is also the last. Oversight that keeps pace with capability is basic prudence.

The Times of Israel

1000 Years of AI Safety

Techno-skeptic narrative

Doomsday odds carry no more rigor than a gut feeling, since a forecast of an event that could never be scored can't be calibrated. Money flowing from a small circle of believers into the groups producing those numbers guarantees extreme answers, and doom makes great marketing for unprofitable labs. The real hazard is human incompetence putting unreliable chatbots in charge of safety-critical systems.