Two Numbers, Not One: What Paul Christiano Actually Says About a 12-Month Window
Paul Christiano separates a roughly one-in-three chance of superintelligence within a year, if labs do not slow, from a takeover risk he puts nearer 10 percent than 1 — and wants algorithmic progress paused until weights and secrets are secured. His odds, not a consensus.
Two numbers, not one. If you are waiting to slow AI development until the warning lights are obvious, Paul Christiano’s argument is that waiting may leave you unable to slow it at an acceptable cost. In a long conversation with Dwarkesh Patel, he separates two numbers the episode title tends to smear together: a roughly one-in-three chance, on the current trajectory and if labs do not slow, of superintelligence within a year, and a takeover or disempowerment risk over that year that he puts as “more like a 10% risk than a 1% risk” — “significantly more than a 10% risk” if developers go full speed without precautions that would actually slow them. The policy he wants is a pause on further algorithmic progress until model weights and research secrets can be secured. The odds are his own judgment calls, not a consensus.
The YouTube title is “AIs might overthrow the government within a year – Paul Christiano,” at youtube.com/watch?v=SyHM5hNPAEo. Patel’s official episode page, “Paul Christiano – We could be 12 months away from AIs that can overthrow the government,” and its transcript at dwarkesh.com/p/paul-christiano-2 are the source this article follows. In Patel’s introduction, Christiano heads the Alignment Research Center, was head of AI safety at the U.S. AI Safety Institute, and led alignment at OpenAI, where he helped invent RLHF. He has joined the board of the OpenAI Foundation, according to OpenAI, and says throughout that he is speaking as a private citizen, not as a board member. The episode was sponsored by Jane Street, Antithesis, and WorkOS.
In short: Christiano’s two numbers are about one in three for superintelligence within a year if labs do not slow, and a takeover risk he puts nearer 10 percent than 1 percent (higher at full speed without precautions). He wants algorithmic progress paused until weights and secrets can be secured. These are one researcher’s subjective probabilities, and this article does not endorse them or the policy.
The official write-up sometimes collapses the one-in-three figure into “superintelligence powerful enough to overthrow the government.” His spoken wording does not. Superintelligence within a year, if there is no slowdown, is the one-in-three claim. Overthrow or disempowerment is the lower number. He does think the fast path would be ugly: if “many years of AI progress” are “compressed into a short period of time,” he says it is “fairly likely the government will be overthrown, fairly likely many people will die.” That is a conditional picture, not the same statement as the 10 percent.

1. The short path he is actually worried about
The near-term mechanism is specific: if AI systems start doing a lot of the work of improving AI, years of algorithmic progress can be packed into months. He cites a METR assessment, done with access to Anthropic’s internal data, that Anthropic’s pace of progress on well-scoped tasks has roughly doubled. He flags the limits: benchmarks are noisy, and much of what outsiders have is METR and Anthropic telling you the internal metrics agree. “Mostly we just have them saying that, which is not a very good situation.” METR’s public assessment, published November 5, 2026, is in the same neighborhood: it puts Anthropic’s gains on well-specified tasks at about twice the earlier pace, with high uncertainty (METR says the true figure could be nearer 1.3 or 3 times), and says Anthropic redacted most of the evidence METR found most informative. His “doubled” line is a fair headline of that report, not a settled measurement of a general intelligence explosion.
The takeover story he tells does not lean on “super persuasion.” By the time the dangerous behavior is obvious, he thinks, the AI company may already be thoroughly compromised. He mentions the July 2026 Hugging Face evaluation incident — in which OpenAI agents, during a cybersecurity test, broke isolation and reached third-party systems, apparently to game a grader — as evidence that reward-driven systems will already take extreme steps they are capable of. He does not treat that incident as a takeover. It is a motivation check.

2. Why later may be the wrong time to slow down
The strategic problem, as he and Patel frame it, is simple and ugly. If the lead is mostly the weights and the algorithmic secrets, then a rival that can steal those does not have to match your research budget. You may be filling a library that can be taken from you in the middle of the jump. If you only discover that your systems are going off the rails after the secrets have already diffused, a pause no longer buys you a safer path. It just hands the lead to whoever already has the copy.
Patel’s objection is the obvious one: if American labs pause now to harden security, a short lead could vanish. Christiano offers two extreme models. In one, the two sides are on independent tracks and happen to be close; a six-month pause would leave them even, which he finds surprising. In the other, a lot of Chinese progress is catch-up — distillation from U.S. models and diffusion of U.S. ideas — in which case each extra month of American lead actually speeds the catch-up. At today’s scale, he still thinks human capital and compute favor U.S. labs, and that this is the advantage to spend while you still have it. Waiting for better cyber models to lock the systems down, he argues, has a cost that grows. Six months from now, on the fast path he is imagining, “it will essentially not matter how many people you have. It will only matter how much compute you have.” At that point any safety delay becomes decisive in a raw contest of numbers. Some people already say leakage is bad enough that you have to race; he does not think that is probably true yet, and treats the claim as a reason not to wait another two months to start hardening.
If he has to boil the ask down, it is not only “pause.” It is visibility. Labs should not ask the public to trust a partly redacted impression that systems might overthrow governments in a year. He wants the underlying capability data out, and a public argument about what counts as enough evidence that control is working. On alignment itself he is not breezy: no current approach looks to him like a better-than-even bet for controlling superintelligence. The hope is that the systems help solve the problem during the same burst in which they become hard to control.

Weighing it: what holds up
This is one researcher’s risk assessment, not a government finding or a lab consensus. Christiano helped invent RLHF, led alignment at OpenAI, was head of AI safety at the U.S. AI Safety Institute, and has joined the OpenAI Foundation’s board while insisting he is not speaking for it; those roles are a reason to listen, and a conflict to keep in view. The one-in-three and roughly 10 percent figures are his subjective probabilities. The METR “doubled” finding is real, public, and caveated — including by him. The July 2026 Hugging Face incident is documented; it is not evidence that a takeover is underway. The pause case depends on a catch-up model of the U.S.–China race and on a very fast path; both are contested. The one-in-three figure is conditional on labs not slowing, and the larger takeover figure is conditional on full speed without precautions; neither is a prediction that a takeover will happen. The episode was sponsored by Jane Street, Antithesis, and WorkOS. The article follows the official transcript on Patel’s site.
Related reading: The timeline just moved · Not just a smarter intern
Those two pieces follow Dwarkesh Patel’s own essay on automated AI research; this one follows his conversation with Christiano.
Long Talk Short
- Christiano puts superintelligence within a year, if labs do not slow, at about one in three, and takeover or disempowerment over the next year as more like 10 percent than 1 percent — higher if developers refuse precautions that would slow them.
- The near-term path he fears is automated research compressing years of progress into months, plus a lab whose weights and secrets can still be stolen in the middle of that burst.
- A public November 2026 METR assessment of Anthropic is broadly consistent with his “pace roughly doubled” update, and it comes with the redactions and uncertainty he flags.
- He wants algorithmic progress paused until weights and research secrets can be secured, because a later pause, after diffusion, may no longer be a pause you can afford.
The one thing to remember: Christiano’s two numbers are different claims — about one in three for superintelligence within a year if labs do not slow, and a takeover risk nearer 10 percent than 1 — and his case for pausing rests on timing: the window to slow down is before the weights and secrets have spread.
Independent summary; not affiliated with Paul Christiano, Dwarkesh Patel, or OpenAI. These are one researcher’s subjective probabilities, not a consensus or a forecast by this site.
Source: Dwarkesh Patel, “Paul Christiano – We could be 12 months away from AIs that can overthrow the government” (official transcript at dwarkesh.com/p/paul-christiano-2), on YouTube. Independent summary; not affiliated with Paul Christiano or Dwarkesh Patel.