The AI Futures Project has moved its conditional timelines sharply forward: all six contributors now put a Superhuman AI Researcher somewhere in 2027, assuming there is no substantial deliberate slowdown in frontier development. That is a forecast, not a report that such a system exists or that its arrival is certain. Its most consequential new input is an independent but partly redacted METR assessment of research acceleration at Anthropic—and the evidentiary limits around that assessment matter as much as the dates.

What changed in the forecast

In its November 10 Q3.5 update, the AI Futures Project says all six authors now expect a Superhuman AI Researcher (SAR) to arrive at some point in 2027, provided companies do not substantially slow or “pace” frontier AI development. The authors also estimate a 35%–65% chance that Top-Expert-Dominating AI (TED-AI)—their category for systems able to automate all white-collar work—arrives before 2028. These are the contributors’ probability judgments, not measured capabilities or a consensus among all AI researchers. The project’s update says the single biggest source of the revision was a new METR report.

The terms are important. SAR is an operationalized milestone: an AI that could fully automate coding in an advanced AI project and match the research taste of a team if its researchers had the research-direction and experiment-selection ability of the best human researcher. “Research taste” here means choosing promising lines of inquiry and interpreting results, not simply producing code. The project describes the threshold as roughly the point when an AI company would prefer automated AI researchers to a team of top human researchers working without AI. TED-AI is a different, broader milestone, and estimates for one should not be casually substituted for the other.

The update compares its implied rate of progress with the group’s AI 2027 scenario, saying the estimated AI-driven acceleration is about five months ahead of that scenario on one modeled measure. This is not a claim that an intelligence explosion is underway on a known countdown. It is a comparison within a forecast framework, conditional on estimates and definitions that remain debatable.

The METR estimate underneath it

METR says it conducted a high-access assessment at Anthropic from August 12 to October 29, 2026. Its team reviewed internal materials and metrics, interviewed staff, and examined model performance and research processes. METR’s subjective conclusion is that, as of October, Anthropic was likely improving model capabilities on well-specified tasks at around twice the rate seen between March 2024 and October 2025. It says it would not be surprised if the true acceleration were closer to 1.3 times or three times the earlier rate. Its empirical fit to Anthropic’s internal AECI benchmark trend estimated a 1.9-times slope after October 2025, with a 95% confidence interval of 1.5–2.8 times.

METR further guessed that a majority—perhaps about 80%—of the current acceleration was attributable to AI use in research and development. The AI Futures Project translated that judgment into an estimated 1.7-times AI-attributable acceleration relative to its historical baseline. That conversion is not a separate direct measurement. METR itself emphasizes substantial uncertainty in attribution, and the public should distinguish the fitted benchmark trend from the more interpretive estimate of what caused it.

METR reports that Anthropic’s training-compute growth did not explain the acceleration it observed. It also says interviews and surveys were consistent with AI use speeding up capability work, while noting that self-reports are challenging to interpret. Importantly, an employee’s reported productivity boost is not the same as a doubling in the pace of overall AI progress: model development also depends on compute, data, algorithms, experiments and diminishing returns. A large uplift for some research tasks does not mechanically produce an equal-sized increase in system-wide capability progress.

How the project translates acceleration into dates

The Q3.5 post does not rely on one route to a year. Its simpler approaches extrapolate different signals and produce different dates. Linear extrapolation of recent research-taste progress points to a median around July 2027 when parameter uncertainty is sampled, or September with median inputs; one particular plug-in calculation elsewhere in the discussion yields November 2027. An extrapolation of staff-reported labor uplift points to around August 2028, but the authors warn it may take self-reports too literally and does not model the effects of automation. Extrapolating the share of R&D tasks at the highest automation level gives roughly November 2027, while a software R&D toy model gives December 2027.

These methods are not interchangeable measurements. They make distinct assumptions about current research taste, the relationship between general capability and research ability, how quickly automation levels advance, and how research effort becomes progress. The project explicitly notes sparse evidence for some parameters. For example, its automation-level extrapolation assumes about seven months between successive automation levels, an assumption it describes as based on little data. Each date should therefore be read as an output of a model rather than an observed schedule.

The project also filters trajectories from its AI Futures Model (AIFM) for acceleration consistent with the new estimate. Conditioning on 1.7-times acceleration in November yields a February 2027 SAR median in one model run; including an assumed 1.2-times AI-attributed acceleration in May moves the results later. The update explains that adding historical conditions changes which model trajectories remain plausible. The spread between results is informative: even when researchers agree the new evidence deserves an update, their estimates depend on how the evidence is translated into model parameters.

In its all-things-considered judgments, each of the six contributors places a median in 2027, with medians spanning May through November. A median is the date by which a forecaster assigns half of their probability mass to the milestone having occurred—not a promise that it will happen then. The project’s forecasts are expressly conditional on no substantial additional slowdown relative to recent pacing. Policy, investment or technical constraints could change the timeline.

What the evidence cannot establish

The most significant public limitation is that Anthropic redacted a majority of the evidence METR considered most informative. METR says some of the redacted material had multiple plausible interpretations and that the public cannot independently inspect key parts of the reasoning. Its report discloses important conclusions, but also says the redactions prevent the ideal of “transparency about evidence.” That means readers are being asked to place substantial trust in METR’s judgment rather than reproduce the entire assessment from published data.

The publicly described AECI trend is useful but imperfect. METR says the post-October 2025 benchmark slope was estimated at 1.9 times the earlier slope, yet cautions this could reflect overfitting to the benchmark distribution or another artifact of the AECI methodology. The data points are relatively sparse and lag research acceleration. Performance on well-specified, verifiable tasks may not map cleanly onto broader, less well-specified scientific or engineering work. Nor can a trend measured inside Anthropic be assumed to describe every AI developer or the industry as a whole.

METR itself lists ways its estimate could be too high or too low. It could overestimate AI’s contribution if non-AI factors, the benchmark’s focus on verifiable tasks, or unusually weak earlier growth have been underweighted. It could underestimate acceleration if growing automation, recent increases in some rates, or benchmark scores that understate real-world impact have been underweighted. The report says it did not identify clear bottlenecks likely to stop further progress in the next year, but immediately qualifies that finding: its understanding changed during the review, and it is not confident it has identified every bottleneck.

AI Futures also acknowledges that its earlier model did not anticipate the observed acceleration well. The post explores several possible explanations, including more R&D value from near-automated coding than assumed, contributions from skills beyond coding and research taste—such as data generation—and a model that translated research effort into progress too conservatively. The authors call this analysis preliminary and say they could not reach strong conclusions. It would be misleading to describe the model’s revision as validation of its prior assumptions; the update is partly an account of unresolved model error.

Why the forecast matters beyond a date

For policymakers and companies, the significance is less whether a particular milestone falls in May or November than whether AI-assisted R&D is becoming a material feedback loop. If models help researchers build more capable models, progress could compound. But the available evidence does not establish that a runaway or self-sustaining “intelligence explosion” is inevitable. The 2027 label summarizes a conditional scenario whose likelihood depends on technical trajectories, human choices, and disputed estimates.

The authors use the update to argue for pausing or pacing frontier development and for an international agreement, on the condition that their timeline assessment is right. That is their policy recommendation, not an empirical conclusion from METR’s measurements. A responsible reading keeps these layers separate: METR’s assessment of one laboratory; AI Futures’ interpretation and forecast; and the policy preferences the forecasters derive from that possibility.

The next useful evidence would include better public methods for tracking AI’s contribution to research, clearer operational tests for autonomy and research quality, and repeated assessments that can show whether the measured acceleration persists beyond one lab and one period. Forecasts can help organize decisions under uncertainty, but their value depends on openly stating what would change them. On the evidence available today, the update is a reason to scrutinize AI-development monitoring and preparedness—not grounds to report a 2027 breakthrough as settled fact.

FAQ

Does AI Futures say a Superhuman AI Researcher will definitely arrive in 2027?

No. The six authors forecast a 2027 median range under a condition: no substantial deliberate slowdown. These are probabilistic estimates, not certainties or observed events.

What did METR actually find?

METR judged that Anthropic’s model capabilities on well-specified tasks were improving at around twice the earlier rate, with substantial uncertainty. It estimated a 1.9-times post-October 2025 AECI trend slope, but much of its most informative evidence was redacted.

Is a 1.7-times acceleration the same as researchers being 1.7 times more productive?

No. The project’s 1.7-times figure is its interpretation of AI-attributable overall progress acceleration. Labor uplift concerns researchers’ task speed and cannot be equated directly with system-level progress.

What should readers watch next?

Look for transparent, repeated measurements, independent replication, broader assessments beyond Anthropic, and evidence about whether acceleration persists. Also track whether definitions of research automation become testable and consistent.

References