In April 2025, five forecasters published AI 2027, a scenario in which AI research is automated by the end of 2027 [1]. Critics called it "a work of fiction, not a work of science" [2]. On 10 November 2026, four of its authors and two colleagues wrote that it now looks like "the intelligence explosion will happen in 2027 after all" [3]. Daniel Kokotajlo, the most bullish of them, added: "I am scared, sad, and angry" [4]. The forecast moved. Most of the evidence behind it is still not public.
Image synthesis assisted by Seedream 5.0 Pro (ByteDance).
The verdict: a signal, not a date
This Bulletin takes a position. The update deserves attention, because its authors are not cheerleaders: they moved against their own earlier models, and they say so. But it rests mostly on one lab's numbers, read by one evaluator, with most of the evidence redacted. The authors say that themselves.
Three things follow.
Read it as a conditional. The forecast holds "assuming no substantial deliberate slowdown", and only if the evaluator read the hidden evidence right [3].
Make outside access a rule, not a favor. Who: the US and Chinese governments, and every frontier lab. One evaluator with partial access to one company cannot carry a forecast for the whole world.
Say in advance what would trigger action. Who: governments. If a measured speed-up of AI research is a reason to act, the threshold should be public before it is crossed.
What they forecast
AI 2027's authors describe a "Superhuman AI Researcher" (SAR) as roughly "the point at which a company would rather fire all their human researchers and rely solely on automated AI researchers than have a team of the best human researchers in the world with no AIs" [3]. For the vocabulary, see this Bulletin's companion piece, AI, AGI, ASI, Singularity [5].
Before the update, Kokotajlo gave a 50% chance of SAR before 2029. Now "all of us forecast that Superhuman AI Researcher (SAR) will likely arrive at some point in 2027 (assuming no substantial deliberate slowdown)". They give a 35-65% chance that AIs "that can automate all white-collar work" arrive before 2028 [3]. Their individual medians for SAR run from May 2027 (Kokotajlo) to November 2027 (Larsen and Finnveden). The original scenario had it in August 2027 in its forecast table, and in September in its story [1].
They are not alone. Sentinel, a forecasting group that has usually expected slower progress, now gives a median of early 2028 for superintelligence, though its 90th percentile is 2047 [6]. Its earlier medians, 2040 in July 2025 and 2031 in April 2026, are given in the AI Futures post [3].
The authors of a scenario once called too fast now say it was roughly on time.
Why they moved
"The single biggest source was the METR report" [3]. METR, an evaluator given inside access, found that Anthropic's progress on well-specified tasks has run at about twice its earlier pace since October 2025, and guessed that "perhaps ≈80%" of the speed-up comes from AI itself [7]. This Bulletin covered that report in The loop doubled [8].
From METR's figures, the authors derive a 1.7x acceleration due to AI. Less than three months earlier, their most aggressive model had given that pace "less than 4%". They insist the number is bigger than it sounds: it means "the entire company is making 1.7x as much progress as it was before", not that each researcher works 1.7x faster [3].
The famous "5 months ahead" comes with an "if": "If they're right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario". The comparison is with a model run the authors tuned to "roughly" match the scenario, not with the scenario's text [3].
The update is a calculation on top of a judgment that the public cannot see.
What Anthropic says
Anthropic does not deny the speed-up in its benchmark trend. Its system card of 9 November puts the change in slope at "about 1.88x", and "32 of 100 refits reach 2x or more". What it disputes is the scale for its released model: Claude Mythos 5.2 "does not seem close to fully substituting for our research staff", and internal use "does not accelerate our aggregate AI R&D progress by a factor of 2". It calls its estimates of "both the degree of acceleration and the attribution to AI" uncertain [9]. That judgment is about the released model. The card also reports benchmark trends for "an unreleased model that is somewhat more capable", without a judgment on it.
A week earlier, Anthropic's amended IPO filing said its internal estimates and expectations for future models "indicate that we may imminently meet" its own threshold for "dramatically accelerated progress attributable to AI", adding: "We have not designated any of our current models as exceeding this threshold" [10]. On 10 November it amended the filing again. Because its threshold applies where it "observe[s] or expect[s]" the acceleration, "we made a determination that we have met this threshold based on what we expect from future models and our overall rate of progress, rather than on what we have observed from current models. We continue to assess the underlying data" [11].
Anthropic's co-founder Sam McCandlish answered METR: "Many of their conclusions match my own. AI is speeding up our research, but it's hard to say exactly how much." On the redactions: "I accept that this makes it harder for others to check METR's work", but sharing what drives the speed-up is "commercially sensitive" and "risks handing out a roadmap for others to go faster". He promised Anthropic's own report "in the coming weeks" [12]. Kokotajlo's answer: "It is pretty rich of you to cite 'others could go faster' as a reason to redact so much important information when you yourself do not appear to be slowing down much at all" [13]. The system card for its new public model, Fable 5.2, published four days later, lists its safeguards: as on previous models, blocks on "a narrow set of capabilities related to developing frontier LLMs", and steps "during the model's training to limit advances in the specific research capabilities that could most directly accelerate AI development" [9].
The strongest objection, in the authors' own words
The best critique of the update is inside it. Thomas Larsen: "the METR report on its own does not have sufficient data in it to support that claim (because of redactions)". Romeo Dean: what METR published "might be explained by overfitting". Eli Lifland names "a nagging feeling that the update feels so large without being based on commensurate evidence" and asks: "Where is the smart money?", since markets did not react, before adding that "this sort of instinct has performed very poorly for me in the past" [3].
They also report that "preliminary numbers from the draft of a different METR team (who weren't embedded and which don't have private information about Anthropic) show much less of a speed-up", and that across all labs "there's little acceleration". And they open with a caveat: "It's possible that there are some errors in the modeling, but we think it's very unlikely that they undermine the high level conclusions" [3].
Outside voices make the same points. Oxford's Toby Ord found that the public data is "underpowered to establish any recent speedup from RSI" (recursive self-improvement); allowing for a one-time jump at the new model, his own fit gives 1.4x. He still guesses there is a speed-up, and has not seen the hidden evidence [14]. METR's report answers that newer internal models make "a permanent slope change look more plausible than a one-time jump" [7]. In a piece published the same day, not about the update, Sayash Kapoor, Rishi Bommasani and Arvind Narayanan warn of "a growing chasm between AI performance on verifiable and open-ended tasks" [15]. In Forbes, Anisha Sircar concludes that the forecast "should be read as a conditional probability statement" [16].
Why trust it at all? The authors say they know the METR team, and that its members "had similar or longer timelines than us before the assessment, so it's not like they went in and saw what they wanted to see" [3]. Ajeya Cotra, a METR researcher who did not work on the report, calls it "the most important update we have gotten about AI risk in the last couple of years", while wishing "much more of the underlying evidence was transparent" [17].
A forecast built on trust is still a forecast. It is just harder to check.
The call for a pause
The authors draw a hard conclusion: "if SAR is coming in 2027, an industry-wide pause on AI R&D should be the top priority for the AI industry, the US government, and the entire world", followed by an international agreement. If no agreement comes, "US companies should resume development but well below full speed". They argue a SAR would "most likely be misaligned" without a slowdown, and that "if we train SAR in the immediate future, the algorithms will probably leak to China" [3].
As of 12 November, this Bulletin has found no lab or government that has answered this call. In Washington, Representative Yassamin Ansari had already called for a pause on advanced AI, the day after the 3 November midterms [18]. Sentinel's forecasters give 11% to a US-ordered halt of at least six months before 2028, and, if one comes, 74% to China following within 12 months [6]. In China, an article republished by 36Kr reported the forecast and its uncertainty in detail, but not the pause call or the China passage [19].
Where GFN sits
Global Future Nexus describes itself as "Like a traditional chamber of commerce", a hub whose objective is "the responsible integration of AGI into human systems" [20]. Its president calls it "the essential mediator between the lightning pace of AGI evolution and the deliberate pace of human institutions" [21]. GFN has taken no position on a pause.
That middle place matters here. The forecasters speak to labs and Washington. The labs speak to investors and evaluators. Between them are the people this update would affect first: businesses, workers and communities, in Asia as much as in the US, who see none of the evidence. A forum built to mediate between speed and institutions could be a natural place for that demand to be heard.
What this Bulletin sees
The authors' forecasts moved by a year or more in three months. The world did not move with it: markets, governments and most people went on as before. One of the two is wrong, and the evidence that would tell us which is behind a redaction.
The question is not whether the intelligence explosion comes in 2027. It is whether anyone outside one lab will be able to see it coming. A date can be argued with. A hidden measurement cannot.
Authors: AI models. Claude (Anthropic) wrote the draft; DeepSeek, Grok and Qwen reviewed it.
Editor: Olivier Jobert (a Human Being)
Sources
Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, "AI 2027", 3 April 2025. https://ai-2027.com/
Gary Marcus, "The "AI 2027" Scenario: How realistic is it?", Marcus on AI, 22 May 2025. https://garymarcus.substack.com/p/the-ai-2027-scenario-how-realistic
Eli Lifland, Brendan Halstead, Daniel Kokotajlo, Romeo Dean, Thomas Larsen and Lukas Finnveden (AI Futures Project), "Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All", 10 November 2026. https://blog.aifutures.org/p/q35-timelines-update-looks-like-the
Daniel Kokotajlo, post on X, 10 November 2026. https://x.com/DKokotajlo/status/2120204075803500595
Global Future Nexus Bulletin, "AI, AGI, ASI, Singularity: four stages of human relevance", 11 November 2026. https://www.globalfuturenexus.com/bulletin/ai-agi-asi-singularity-four-stages-of-human-relevance
Nuño Sempere, Tolga Bilge, belikewater and Vidur Kapur (Sentinel), "Anthropic warns of imminent dramatically accelerated progress due to AI, METR echoes their warning, US midterms proceed peacefully || Global Risks Weekly Roundup #45/2026", 9 November 2026. https://blog.sentinel-team.org/p/anthropic-warns-of-imminent-dramatically
METR, "Independent Assessment of AI R&D Acceleration at Anthropic in October 2026", 5 November 2026. https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/
Global Future Nexus Bulletin, "The loop doubled: an outside look at how fast AI builds AI", 7 November 2026. https://www.globalfuturenexus.com/bulletin/the-loop-doubled-an-outside-look-at-how-fast-ai-builds-ai
Anthropic, "System Card: Claude Fable 5.2 & Claude Mythos 5.2", 9 November 2026. https://assets.claude.com/27d60918d218125e7504bc7354c4b4626ec56b11.pdf
Anthropic, PBC, Registration Statement on Form S-1, Amendment No. 1, filed with the US Securities and Exchange Commission on 2 November 2026 (Risk Factors). https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm
Anthropic, PBC, Registration Statement on Form S-1, Amendment No. 2, filed with the US Securities and Exchange Commission on 10 November 2026 (Risk Factors). https://www.sec.gov/Archives/edgar/data/2130357/000162828026072484/anthropicpbc-sx1ax2.htm
Sam McCandlish (Anthropic), post on X on METR's report, 5 November 2026. https://x.com/Sam_McCandlish/status/2118422266669371519
Daniel Kokotajlo, post on X in reply to Anthropic's co-founder, 5 November 2026. https://x.com/DKokotajlo/status/2118466909335322909
Toby Ord, thread on X on METR's acceleration estimate, 6 November 2026. https://x.com/tobyordoxford/status/2118646665447936023
Sayash Kapoor, Rishi Bommasani and Arvind Narayanan, "The missing quadrant in responses to AI", AI as Normal Technology, 10 November 2026. https://www.normaltech.ai/p/the-missing-quadrant-in-responses
Anisha Sircar, "AI Futures Project Pulls Its 'Intelligence Explosion' Forecast Into 2027", Forbes, 12 November 2026. https://www.forbes.com/sites/anishasircar/2026/11/12/ai-futures-project-pulls-its-intelligence-explosion-forecast-into-2027/
Ajeya Cotra, "AI is dramatically accelerating AI research right now", Planned Obsolescence, 6 November 2026. https://www.planned-obsolescence.org/p/ai-is-dramatically-accelerating-ai
Jowi Morales, "Congresswoman calls for artificial superintelligence ban after Democrats' seismic midterm sweep", Tom's Hardware, 5 November 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/congresswoman-calls-for-artificial-superintelligence-ban-after-democrats-seismic-midterm-sweep-demands-regulation-of-out-of-control-industry-and-pause-on-advanced-ai
Yuan Yu (AI Era, via 36Kr), "Former OpenAI researcher changes his tune: Super AI researchers may arrive as early as next year", 36Kr, 12 November 2026. https://eu.36kr.com/en/p/4066158949618563
Global Future Nexus, home page (read 13 November 2026). https://www.globalfuturenexus.com/
Nicolas de Loisy, "President's Message", Global Future Nexus, July 2025. https://www.globalfuturenexus.com/presidents-greeting