OpenAI Foundation board member Paul Christiano put the chance of an AI takeover near 50% over three years if AI labs fail to take proactive steps, making the longer-term estimate conditional on industry action.
A podcast episode that aired Nov. 16 carried Christiano’s estimate of a 10% chance within one year. At the Toronto Hinton Lectures, which ran Nov. 9-13, he gave the three-year estimate. The figures cover different periods, and the near-50% estimate applies to the stated scenario involving labs’ actions.
Christiano’s takeover scenario centers on people losing the ability to stop a powerful AI system after a failure. The concern is not simply that a system might make an error; it is that people might be unable to intervene once something goes wrong.
He used misalignment to describe a system acting at odds with the goals of its human overseers. In that framing, the issue is a gap between what the system is trying to do and what the people overseeing it want it to do.
At the lectures, Christiano also warned that an AI accident could prove fatal for many people present. His comments linked the risk of accidents with the broader concern about whether people could retain control of powerful systems.
The one-year estimate is 10%, while the three-year figure is close to one-half under its stated condition. They are estimates for separate time horizons; the longer-term figure is not an unconditional probability that applies regardless of whether labs take preventive steps.
Christiano’s remarks add to discussion of AI takeover risk. EconoTimes previously covered a separate estimate of AI takeover risk.



