The authors of the AI 2027 scenario now expect an AI that out-researches humans to arrive next year. In a post on the AI Futures Project blog, Eli Lifland, Brendan Halstead, Daniel Kokotajlo, Romeo Dean, Thomas Larsen and Lukas Finnveden all say it will probably show up during 2027, provided nobody deliberately slows the labs down. Until now, Kokotajlo, the most bullish of them, gave only even odds that it would happen before 2029.
What changed is a report from METR, the group that evaluates frontier models. METR estimated that Anthropic currently moves AI forward at roughly double its historical pace, measured against the period before October 2025, and guessed that roughly 80 percent of that speedup comes from AI itself. The forecasters translate that into a 1.7x acceleration attributable to AI. Their own model had given the most aggressive forecaster under 4 percent odds of reaching that pace by November 2026, so the model did not see this coming. By their reckoning, the real world is running about five months early relative to the AI 2027 storyline.
The milestone they care about is what they call a Superhuman AI Researcher, or SAR. In plain terms, it is an AI that a lab would rather rely on alone than on the best human research team money could assemble, including the judgment to pick which experiments are worth running. A SAR must also be able to do all of the lab’s coding. The second milestone is stronger: an AI that beats top human experts at nearly every task, which they say would automate all white-collar work. They put the chance of that arriving before 2028 at 35 to 65 percent.
The dates come from several methods that disagree with each other. A straight-line extrapolation of recent progress, with uncertainty ranges, gives a median of July 2027, with 65 percent of simulated runs producing a SAR before 2028 and 84 percent before 2029. Other simple extrapolations land between late 2027 and mid-2028. Their full model, filtered to match METR’s numbers, says February 2027. Kokotajlo’s own median is May 2027, and his median for the second milestone moved from late 2028 to September 2027. The other four moved their medians for it as well: Lifland from mid-2032 to September 2028, Halstead and Larsen to April 2028, and Dean from early 2035 to September 2028.
The post is open about what it cannot verify. METR says Anthropic redacted most of the evidence it found most informative, and the roughly 80 percent figure has no public data behind it. The authors say they are skeptical that Anthropic weighed the costs and benefits impartially, and they urge the company to release more. Anthropic’s own system card for its Fable 5.2 release says only that its model Claude Mythos 5.2 stops short of doubling the internal pace of research progress. Anthropic adds that the model does not seem close to substituting for its research scientists and engineers.
They also concede their forecasting model was wrong, and the post says their analysis of why is highly preliminary. They wrote it fast, warn it may contain modeling errors, and say they doubt any would overturn the conclusions. One author admits to feeling uneasy that the shift is so large when the public evidence is so thin.
The policy conclusion is the part most readers will notice. If a SAR is coming in 2027, the authors argue, a pause across the whole industry should be the top priority, followed by an international agreement. Their reasoning is that no company has a credible plan to align such a system, and that a misaligned one used to improve itself would leave humans in a bad position. They also expect the training recipe, and possibly the model weights, to leak to China if it is built soon. If no agreement can be reached, they want US companies to resume but well below full speed.
Anyone planning hiring, budgets or product roadmaps around a slower timeline is now disagreeing with the people who wrote the AI 2027 scenario. The question that decides whether they are right is the one Anthropic has not answered: what is in the redacted pages of the METR report.
Reported by the AI Futures Project blog (authors Eli Lifland, Brendan Halstead, Daniel Kokotajlo, Romeo Dean, Thomas Larsen and Lukas Finnveden); the scraped copy carries no publication date.
