The AI 2027 scenario depends on a conjunction of individually improbable developments.
4 events · 1 assessment · 2 decisions
Linked subclaims
After matching, linked two existing claims as contradicting considerations: superhuman coding ability by 2027 (6c1f5ff8) and AI automation of AI research triggering rapid recursive growth (e68a817d). These are the two cruxes named in the assessment. No named arguments were created, and the assessment stands as recorded. The claim's one instance was not reviewed this pass.
Assessed first pass
Budget was very constrained ($0.40 used in the first turn). The claim was assessed as inconclusive at 0.55 confidence, with no credence because it is evaluative and composite. It was left atomic: subclaims on AI timelines and takeoff speed were identified in the trace but not minted, because there was no budget for matching. Importance was set to 0.4 and contestation to 0.6. Instances were not reviewed this pass. A future pass should add the timelines and takeoff subclaims via match_claim and review the claim's one instance.
Assessed Inconclusive
verdict confidence 0.55
AI 2027, published in April 2025 by Daniel Kokotajlo, Scott Alexander and colleagues at the AI Futures Project, is a month-by-month narrative of AI progress through 2027. In that story a superhuman AI coder arrives, AI research automates itself, and the world reaches one of two endings. Part of this claim is clearly true. Any narrative that detailed is a conjunction, and the specific path it describes is unlikely as a whole. The authors say as much: they present 2027 as roughly their modal year, not their median, and their published forecasts put most of their probability on later dates. Critics, including commentators who reconstructed the scenario's timelines model, argue that its key inputs are thinly supported and that several steps each need fast progress to happen on schedule. The disputed part is the word "individually." The authors' own forecasts give the first step, a superhuman coder by about 2027, well under even odds. They treat the later steps as fairly likely once that step happens, because faster AI research feeds on itself. So whether each link is improbable depends mostly on two contested questions: how short AI timelines are, and whether automating research produces a fast takeoff. Critics put low odds on several links. Proponents say the steps are strongly correlated, so multiplying separate probabilities understates the chance of the overall picture, even if the exact sequence is unlikely. The question would be settled by evidence on the near-term trend in AI coding ability, and on how much automating AI research actually speeds it up. Evidence through late 2026, including the authors moving their own timelines later, fits the view that the scenario's pace on its own dates is improbable. It does not settle whether the underlying steps are individually improbable.
Claim entered the graph