With the new year just around the corner, there are plenty of questions around what 2027 could hold for artificial intelligence.

Interestingly, that is the year the AI Futures Project — the forecasting group behind the widely discussed AI 2027 scenario published in April 2025 — now says there is a meaningful probability that a “Superhuman AI Researcher” will likely emerge. Earlier this week, the group moved its timeline forward, citing an independent evaluation of Anthropic’s internal research practices as the biggest source of new evidence.

The revision is an interesting stress test for how to think about forecasting, as well as about the reliability of the institutions producing these forecasts.

The original AI 2027 scenario depicted a rapid takeoff in which AI systems begin improving themselves, compressing years of research progress into months. However, by August 2026, most of the authors had pushed their median dates for “Top-Expert-Dominating AI” into the 2030s. As of January 27, 2026, Daniel Kokotajlo’s superhuman coder median was at roughly end-2029 or early 2030. By August 2026, his median for Top-Expert-Dominating AI was late 2028, while Eli Lifland’s was mid-2032.

The catalyst for the reversal was a report published November 5, 2026, by METR , a nonprofit AI evaluation organization. METR had spent eleven weeks examining data including Anthropic’s internal research materials, model performance data and employee interviews. Most METR staff on the project concluded that as of October 2026, Anthropic was likely improving its models on well-specified tasks at around twice the rate it had between March 2024 and October 2025. METR guessed that perhaps 80% of the current acceleration was due to Anthropic’s internal use of AI, though Anthropic redacted most of the informative evidence METR relied upon.

METR’s estimate came with a range: it said it would not be surprised if the true acceleration were closer to 1.3x to 2.8x faster. In a separate review of Anthropic’s August Risk Report, it put leading-indicator estimates between 1.3x and 2.8x and cited “substantial methodological uncertainty” about how to measure them. In an amended S-1 filed three days before the METR report, Anthropic said its internal estimates and expectations for future models mean it “may imminently meet” its safety policy’s definition of “dramatically accelerated progress attributable to AI”. It said it has not designated any current model as exceeding that threshold.

The AI Futures Project calculated that this implied an AI-driven speed-up of roughly 1.7x in overall AI research progress. Even the most aggressive of the group’s previous model setups, Kokotajlo’s, had assigned less than a 4% probability to this level of acceleration by November 2026.

‘Scared, sad and angry’: The New Forecast

The six researchers behind the update — Eli Lifland, Brendan Halstead, Daniel Kokotajlo, Romeo Dean, Thomas Larsen and Lukas Finnveden — now all forecast that a Superhuman AI Researcher will likely arrive in 2027, assuming no substantial deliberate slowdown. Their individual medians fall between May and November 2027 . Kokotajlo assigned a 50% chance by May 2027. The others’ medians come later in 2027, including August for Lifland and Dean. They define the milestone as an AI that matches the research taste of the best human researchers at an AGI company and can fully automate its coding, roughly the point at which a company would rather rely on it than on the best human researchers working without AI.

On X , Kokotajlo wrote: “Unfortunately, it seems like the intelligence explosion will be happening in 2027 after all. If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario. I am scared, sad, and angry.”

The authors also put the chance that “Top-Expert-Dominating AI,” which they describe as AI that can automate all white collar work, arrives before 2028 at between 35% and 65% .

The METR findings describe a feedback loop inside the companies that build consumer AI products. If AI is accelerating the research that produces the next generation of AI, then the capabilities available to consumers — in writing assistants, coding tools, search and creative software — may improve faster than the product release cycles that consumers have learned to expect.

That is no longer purely hypothetical: METR found that AI is already driving much of the acceleration at Anthropic. In its 11 week review from August to October 2026, it did not identify a compute, data or algorithmic bottleneck likely to stop the acceleration from continuing. It added that it was not confident in that assessment and that Anthropic redacted its analysis of specific bottlenecks .

The forecast is also a warning about the limits of consumer trust infrastructure. During the same that week METR published its report, Anthropic filed filed an amended S-1 ahead of its planned IPO, saying it may imminently meet its threshold for accelerated progress with future models. Consumers are being asked to trust companies that are themselves uncertain about the pace of their own systems. RAND’s March 2026 assessment of AGI forecasting concluded that the field “lacks resolved forecasts for calibration” and that decision-makers are “making decisions based on methodologies that are in nascent stages of development”.

Furthermore, the forecast compresses the timeline for the labor market disruption that has been the most tangible consumer concern about AI. The AI Futures Project’s estimate that AI capable of automating most office work could arrive before 2028 is a signal that the buffer between “AI tools that assist workers” and “AI systems that replace tasks” may be shorter than mainstream economic forecasts assume.

The AI Futures Project update is not the first time experts have moved AI timelines dramatically. It is part of a pattern that should temper confidence in any single forecast.

The AI Impacts survey of machine learning researchers found a median expectation of human-level machine intelligence by 2047 in 2023, compared with 2060 in 2022 — a shift of thirteen years in a single year. The same survey estimated a 10% chance of unaided machines outperforming humans in every task by 2027, and a 50% chance by 2047. RAND’s 2026 synthesis found that AGI timeline estimates have shifted earlier across independent methods, but that “all methods share significant limitations”.

The historical record of AI prediction is worse. A 2014 study of famous AI forecasts — from the 1958 prediction that a computer would be chess champion within a decade to later projections — concluded that “the general reliability of expert judgement in AI timeline predictions is shown to be poor”. Newell and Simon’s chess prediction took nearly four decades rather than ten; Deep Blue beat Garry Kasparov in 1997.

The AI Futures Project’s authors have revised their dates repeatedly, in both directions. As of August 2026, Kokotajlo’s median for a Superhuman AI Researcher was December 2028. By November 2026, it was May 2027. The group’s February 2026 grading of AI 2027 found quantitative progress running at roughly 65% of the predicted pace, though most qualitative predictions were on track. A preliminary July update put the figure nearer 75%. The latest METR-based estimate, the group says, puts AI-driven acceleration about five months ahead of the AI 2027 scenario.

The forecast, then, should be read as a conditional probability statement. The group floats it that way, too: if a Superhuman AI Researcher is coming in 2027, it says , an industry-wide pause on AI R&D “should be the top priority for the AI industry, the US government, and the entire world.” That is a policy recommendation contingent on the forecast.

A few tensions will likely determine whether the 2027 forecast proves prescient or joins the long list of AI predictions that did not survive contact with reality.

METR’s most informative evidence was redacted by Anthropic. The organization’s estimate of 2x acceleration is based on partial access. The range of roughly 1.3x to 2.8x is wide enough to accommodate both a modest productivity gain and a true inflection point. Independent verification is still very limited.

Also, “Superhuman AI Researcher” is a constructed category. The AI Futures Project defines it as an AI that matches the research taste of the best human researchers and can fully automate coding. However, “fully automates” and “conducts R&D” are not binary states; the transition from human-in-the-loop assistance to autonomous research is likely to be contested and uneven across domains. A date on a calendar may obscure more than it reveals.

Finally, RAND’s assessment is that the field lacks benchmarks resistant to saturation and gaming, continuous real-time insight into model capabilities and independent validation of influential models. RAND calls the degree to which AI accelerates AI research the leading indicator most relevant to rapid capability gains. It notes that frontier laboratories are best positioned to track it internally and should “develop and improve standardized monitoring systems now”. The AI Futures Project’s update rests chiefly on one lab’s data, evaluated by one nonprofit, with substantial redactions. The acceleration also appears specific to Anthropic: METR’s all-developer benchmark trend shows about 1.4x, and Anthropic says its current models do not double the pace of its research.

The next twelve months are therefore likely to provide tests that are more concrete than any forecast.

METR said it “cannot rule out a substantial further acceleration” within the next six months. Anthropic’s IPO filing will subject its internal claims to a level of public scrutiny that private companies do not witness. The AI Futures Project has been updating its timelines roughly every quarter, and each new update will either narrow or widen the gap between forecast and reality.

The practical question is less about whether a Superhuman AI Researcher arrives in May or November 2027 — it is more whether the products consumers use, the tech environment, the jobs and the institutions are adapting at a pace that matches the acceleration METR measured. On that question, the forecast is a warning, but the evidence is still being collected, and the verdict is still wide open.