Paul Christiano, who recently joined OpenAI's board, says AI labs should pause efficiency gains in training until their security is adequate and they can tell whether they control their systems

Dwarkesh Podcast · Video interview

Read the source (opens in a new tab)
What happened
Paul Christiano, head of the Alignment Research Center, who "recently joined OpenAI's board," discussed near-term AI risk in a Dwarkesh Podcast interview published November 16, 2026.
Key points
  • He put the risk of AI takeover over the next year at "more like a 10% risk than a 1% risk," and "significantly more" if developers move "full speed ahead" without precautions.
  • He gave "maybe a 1/3 chance that we have superintelligence within one year on the current development trajectory, if we do not slow."
  • He said developers should "pause further algorithmic progress on those problems," meaning more efficient pre-training, reinforcement learning and architectures, until "their security posture is adequate" and they can tell "whether they have adequate control of the systems they've built."
  • He called for "more public visibility into the capability trajectory" and said the ability to stop if things go badly "is not where we are now."
Evidence
Podcast interview, read from the publisher's own transcript; the video was not watched. Christiano gives his own estimates, says he is "not speaking as an OpenAI board member," and wants to discuss "the limitations of what we actually know."

See it in the latest record