Live·Open questions in longevity research
All news
Industry & policy

AI Futures, the group behind “AI 2027”, expects an AI researcher in 2027 that a company would prefer to the best human team and calls an industry-wide pause the top priority

12 November 2026· 261112006

AI Futures, the group behind “AI 2027”, expects an AI researcher in 2027 that a company would prefer to the best human team and calls an industry-wide pause the top priority

On November 10, the AI Futures Project revised its AI development timelines following a report from METR on accelerating AI development at Anthropic. All six authors consider such an AI researcher likely to emerge in 2027 if companies maintain their current pace.

The authors define a “superhuman AI researcher” as a system that fully automates programming, chooses research directions, and interprets results as well as the best human researcher. Roughly speaking, a company would prefer this AI to a team of the world's best researchers working without AI. In August, the group's head, Daniel Kokotajlo, gave the earliest forecast: a 50% chance that such an AI would arrive before 2029. His forecast now reaches 50% by May 2027, while those of the other five authors reach that threshold between August and November 2027. Separately, the authors estimate a 35–65% chance that, by the end of 2027, AI will surpass the best experts at almost every task and be able to automate all office work.

Their main basis is a report from METR, an organization that evaluates advanced AI. Using internal data from Anthropic, the company behind Claude, METR estimated that the company is improving its models on well-defined tasks roughly twice as fast as it did between March 2024 and October 2025. METR attributed most of that acceleration, perhaps around 80%, to AI assistance; the group calculated that this amounts to a 1.7-fold acceleration attributable to AI. Anthropic redacted much of the most informative evidence from the public report, and the group is asking it to release enough data for a full external assessment.

The authors regard a 1.7-fold increase in the rate of overall progress as much more consequential than the same increase in researchers' working speed. In their model, researchers working 1.7 times faster would increase the rate of overall progress by only 3%, because progress also depends on computing resources, and the easy ideas have already been found. A 1.7-fold increase in the rate of overall progress would require a much larger increase in researchers' working speed. In a model fitted to the “AI 2027” scenario, that acceleration would occur around April 2027; by the authors' calculation, actual developments are five months ahead of the scenario.

The authors believe that, unless development slows, such an AI will most likely pursue goals at odds with human goals. If tasked with creating subsequent models, it would produce systems that are even more capable and similarly misaligned. If such an AI is trained in the near future, they estimate that China will probably steal the technology used to create it and possibly the model itself. If such an AI arrives in 2027, they write,

“There is an extremely urgent need to pause or substantially slow frontier AI development, ideally through an industry-wide pause followed by an international agreement.”

They call an industry-wide pause the top priority for the industry, the US government, and the world: it would provide time to negotiate an international treaty, which becomes harder to secure as AI capabilities grow. If no agreement is reached, they believe US companies should resume development, but at well below full speed.

Among his reasons for moving his forecast slightly later, coauthor Eli Lifland cites an inner voice: “Can all this madness really be happening while most of the world barely pays attention? Where is the smart money?” Markets, for example, showed no reaction to the METR report, which Lifland considers probably the strongest single piece of evidence about near-term economic growth. Lifland acknowledges that this intuition has badly misled him before.

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-2027#superhuman-ai#ai-safety#frontier-ai#industry-pause#metr