The SURA debate turns on how doomed Plan C is
In late October 2026, Scott Alexander published Against SURA, arguing that the AI safety community has swung too far toward demanding that safety-minded labs and people simply stop. SURA is his name for “Stop Unilaterally + Raise Awareness”. Zvi Mowshowitz replied two days later that the post was “deeply confused”, and hundreds of comments followed. The source page summarizes the post, the replies, and the comment threads.
Read closely, the two sides agree on more than their titles suggest. Most of the remaining disagreement comes down to one number: the chance that a short, rushed alignment effort by a leading lab would work.
What almost everyone agrees on
Using the vocabulary of Plans for the AI transition, where Plan A is a coordinated global pause and Plan C is a careful leader spending its few months of lead on alignment:
- A coordinated global pause is best if achievable.
- Plan A probably will not happen, so preparing for Plan C has some value. Zvi grants this, Scott’s strongest point.
- A coordinated stop beats a unilateral one.
- Current labs could do much more to pursue a pause than they are doing. Scott largely conceded this to commenters who pointed to middle options between disbanding and racing.
Four separate disagreements
- How likely is a rushed alignment effort to work? This is the central one. At the 90% Scott floated on social media, staying in the race to execute Plan C well is clearly right. At Yudkowsky’s near zero, stopping regardless is right, as Zvi concedes. See How hard is alignment and Can AI automate alignment research.
- Does stopping first make a pause more or less likely? Scott sees a lost bargaining chip; others see a costly signal with historical precedent in biological weapons and CFCs. See Can AI progress be paused and Can the US and China agree to limit AI.
- Do safety-minded insiders change what labs do? Scott points to political spending and credibility; Daniel Kokotajlo and TsviBT point to the lack of serious effort toward a pause. See Should safety-focused labs keep building.
- Is building something that might kill everyone forbidden in itself? This is an ethical question rather than a factual one, and no forecast will settle it.
Assessment
Scott’s strongest point is that the backup plan deserves serious preparation because the first-best plan probably fails. His weakest is the framing: setting “disband” against “race” leaves out the middle options that most of his critics actually hold. His AI-forecaster numbers are too fragile to bear the weight he puts on them, as he partly acknowledges in footnotes. Zvi’s strongest point is the conditional: the right strategy depends on the probability that Plan C works, and Scott’s optimism about that probability is where he is least supported.
Aaron Scher’s comment sharpens this. Much of the estimated survival under Plan C is luck about how hard alignment turns out to be, which no one’s effort changes, while survival under a pause depends more on choices people can influence. Comparing raw probabilities may therefore understate the value of working toward a pause.
What remains unresolved
The most useful outcome of the exchange is a sharper question: what evidence, available before the critical moment, would tell us whether a lead of a few months is enough? Candidates include whether alignment techniques keep working across model generations, and whether AI-produced alignment research can be independently verified.