AI Futures moves its forecast for superhuman AI researchers to 2027

The AI Futures authors now expect superhuman AI researchers in 2027, bringing forward their forecasts after an outside assessment of Anthropic. Their November 10 update calls for a frontier-AI pause, but its biggest input comes with a constraint: Anthropic withheld much of the evidence behind the evaluator’s conclusions. The forecasts assume no substantial deliberate slowdown.

The shift is substantial even by the group’s own recent expectations. Less than three months earlier, Daniel Kokotajlo, its most bullish forecaster, had assigned a 50% chance of reaching that milestone before 2029. Now all the update’s authors place its likely arrival in 2027.

Replacing a research team is the milestone

A Superhuman AI Researcher, in the authors’ definition, is not simply a chatbot that writes better code. It marks the point where a company would prefer automated AI researchers alone over a team of the world’s best human researchers working without AI. That is a forecast about replacing the research function, not just speeding up individual assignments.

The update also puts a 35–65% chance on AI capable of automating all white-collar work arriving before 2028, conditional on projects and companies not slowing down. This is the authors’ broader capability forecast. It is not a claim that workplaces will have adopted such systems or replaced their workers by that date.

The Anthropic evidence behind the revision

METR’s November 5 assessment followed high-access work inside Anthropic from August 12 through October 29. Evaluators examined internal model benchmarks, training-compute data, AI speedup analyses and staff interviews. METR described the access as unprecedented, while limiting its headline conclusion to improvements on well-specified tasks.

Anthropic’s internal benchmark index, AECI, showed scores rising at an estimated 1.9 times their pre-October 2025 rate. The statistical confidence interval was 1.5–2.8 times, and different methodological choices produced estimates from 1.1 to 2.9 times. METR warned that benchmark design or overfitting could help explain the result, and that broader, less clearly defined tasks carry additional uncertainty.

Explaining the acceleration is harder than measuring it. METR ruled out faster growth in training compute as the cause, but found no decisive evidence separating all the other contributors. Its judgment that perhaps 80% of the current acceleration came from AI was explicitly subjective and substantially uncertain. Non-AI factors likely mattered, particularly in late 2025.

AI Futures translates those estimates into roughly 1.7 times the overall pace of progress attributable to AI. That is different from making researchers 1.7 times more productive: faster individual work still depends on computing resources and faces diminishing returns. The authors’ concern is a feedback loop in which better AI speeds research, producing still better AI and further acceleration.

From faster benchmarks to a 2027 forecast

The update uses several forecasting methods rather than one countdown. Their different outputs show how assumptions about continued automation affect the projected arrival:

• A simple extrapolation of research judgment and the recent progress rate puts superhuman AI researchers around July or September 2027, depending on how parameter uncertainty is handled. It does not assume additional acceleration from automation.
• An extrapolation of researchers’ self-reported productivity gains points to mid-2028. The authors note that those self-reports may be too optimistic, even though the method does not model further acceleration from automation.
• Filtering the existing AI Futures Model for a 1.7-times AI-attributed acceleration produces a February 2027 median. Adding an earlier acceleration estimate from May 2026 pushes that result later, illustrating its sensitivity to the trajectory being fitted.

The authors call their analysis of why their model missed the acceleration highly preliminary. Possible explanations include greater-than-expected gains from coding, contributions from skills such as data generation, or overly conservative assumptions about how research effort translates into progress. They did not reach strong conclusions about which explanation dominates.

Public scrutiny faces a separate obstacle. Anthropic redacted a majority of the evidence METR found most informative for its conclusions. The company said the redactions protected confidential information that could accelerate competitors. METR said much of that evidence had multiple plausible interpretations, leaving the public more dependent on its judgment than it considered ideal. AI Futures urges Anthropic to disclose enough for outsiders to assess the findings.

METR did not identify clear bottlenecks that would prevent further acceleration over the next year, but its understanding changed repeatedly during the assessment. It remains uncertain about those limits. It also found that current gains primarily came from AI assisting researchers, rather than fully autonomous contributions—a material distinction from the replacement milestone AI Futures forecasts.

For the forecasters, the shorter timeline warrants action despite that uncertainty. They advocate pausing or substantially slowing frontier development, ideally through an industry-wide pause followed by an international agreement. Their argument is that companies are not on track to make superhuman AI researchers reliably aligned with human intentions before using them to improve subsequent systems.