AI Futures shortens its timelines: a superhuman AI researcher likely in 2027

The post is titled Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All. The authors, including Daniel Kokotajlo, Eli Lifland, Brendan and Lukas Finnveden, build models to forecast when AI will reach key milestones. They define a Superhuman AI Researcher (SAR) as a system that can automate coding entirely and has research taste (choosing directions, selecting experiments and interpreting results) at or above the level of the top human researcher. At that point, a lab gains more from spinning up another AI researcher than from hiring top human scientists.

Every forecaster in the project now places their median arrival date for SAR in 2027. Kokotajlo forecasts May 2027, Brendan March, Lifland August and Finnveden late 2027. They also estimate a 35 to 65 percent chance that Top-Expert-Dominating AI (TED-AI), capable of automating all white-collar work, arrives before 2028.

The trigger is a recent assessment of Anthropic by evaluation group METR. METR estimated that AI research progress there is advancing roughly twice as fast as its pre-October 2025 trend, with about 80 percent of the gain downstream of AI. That implies a 1.7x pace-of-progress acceleration attributable to AI alone. AI Futures notes that a 1.7x overall speedup is a much bigger deal than an employee working 1.7x faster: because progress also depends on compute and faces diminishing returns, a 1.7x overall acceleration requires AI to multiply research labor by a large factor. In an informal poll at Anthropic, researchers estimated they were working roughly twelve times faster with AI.

In AI Futures' models, the current pace is roughly five months ahead of the earlier AI 2027 scenario. If a superhuman researcher arrives next year, the authors argue it is urgent to pause or pace frontier AI development, ideally through an industry-wide pause followed by an international agreement. Companies are not on track to align such a system, they write, and recursive self-improvement with a misaligned model poses extreme risks. In addition, the training recipe could be stolen.

In late September 2026, twenty-two researchers, including three Turing Award winners and senior leaders at OpenAI, Anthropic and Microsoft, published a paper on the risks of an intelligence explosion. It warned that automated research could compress years of progress into months.

Two months earlier, more than 1,300 lab employees signed Pacing the Frontier, asking for tools to slow down development.

METR explicitly notes that Anthropic redacted a majority of the evidence behind its conclusions. External observers cannot independently verify what the 2x acceleration and 80 percent AI attribution estimates rest on.

Anthropic itself qualified the figures in its Claude Fable 5.2 system card. The company stated that its internal model Claude Mythos 5.2 does not double the overall rate of research progress, though it acknowledged its measurements were uncertain and promised a future report.

The concept of research taste is an abstraction in the authors' mathematical model, not a fixed physical quantity. Capabilities can develop unevenly, moving rapidly in mathematics and code while lagging on long-horizon projects.