The AI 2027 scenario underestimates the time remaining before general intelligence arrives.
4 events · 1 assessment · 2 decisions
Decomposed
I added two arguments. The argument for the claim rests on a new subclaim that the authors' median was after 2027 (seeded at 0.85, importance 0.3). The argument against it uses the existing METR task-horizon doubling-time quantity. I also linked the existing claim that superhuman coding arrives by 2027 as a contradicting consideration. Both arguments are rated slight. The assessment (supported, credence 0.75) stands. The claim has no dependents to notify.
Assessed without decomposition
Budget allowed only one substantive turn, so I recorded an assessment of supported (credence 0.75, confidence 0.6) from background knowledge and set importance to 0.6 and contestation to 0.7. I made no decomposition and no instance review. A future pass should decompose into the authors' own median versus mode, benchmark trend evidence (METR time horizons), and the definition of general intelligence, and should link to existing AGI-timeline claims.
Assessed Supported
verdict confidence 0.60 · credence 0.75
AI 2027, published in April 2025 by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, describes a superhuman AI coder arriving around early 2027 and superintelligence by late 2027. The claim is that this timeline is too short: general intelligence will take longer than the scenario shows. The evidence leans toward the claim, but the question cannot be settled until the scenario's dates arrive. Three considerations favor it. First, the authors described 2027 as roughly their modal year, not their median. Their own median forecasts for a superhuman coder were later at publication, so by their own lights the scenario sits on the early side of their distribution. Second, in later updates the authors and the AI Futures Project moved their median timelines further out. Third, as of late 2026, AI progress has been fast but has not clearly matched the scenario's intermediate milestones, such as AI systems substantially automating frontier AI research. Critics have also argued that the model's extrapolation of the length of tasks AI can complete, and its assumptions about speedups to AI research, are too aggressive. Against the claim, progress on coding and agentic benchmarks has stayed rapid. The scenario was always offered as one plausible path, and a 2027 to 2028 arrival is not ruled out. The claim would be overturned if systems that automate most AI research appear by about 2028. The answer also depends on how "general intelligence" is defined, which the scenario operationalizes through a superhuman coder and superhuman AI researcher milestones.
Claim entered the graph