Will AI 2027 happen? Prediction-by-prediction tracker

The tracker

Status key: Happened · On track · Behind · Unclear · Not yet due (see How we grade). All rows checked 2026-10-15.

Is AI 2027 realistic? What outside critics say

Short answer. Critics say the scenario is a vivid story resting on a forecast model with real flaws, and that 2027 was never likely. The authors accept some of the model criticism, fixed bugs and moved their own dates later, then somewhat earlier again in 2026. They also point out that 2027 was their most likely year, not their median.3 Whether AI 2027 is "accurate" so far is what the tracker above measures, forecast by forecast.

The critics

"A work of fiction, not a work of science" (Gary Marcus). The cognitive scientist and AI sceptic called the scenario "a fun read" that is "crafted to look like science". He wrote: "I honestly wish, though, that it wasn't being taken so seriously. It's a work of fiction, not a work of science."5 His main argument is that the story is a chain of bets that all have to come in: "pretty much the whole thing is a house of improbable longshots. If any one of those longshots fails to arrive on time, the entire timeline is moved back".5 He also argued that such scenarios are "practically marketing materials for companies like OpenAI and Anthropic, who want you to believe that AGI is imminent".5 He did say he "mostly salute[s] AI 2027's intention", which is to get people to act on AI risk.5

"I think it's pretty bad" (titotal, on the timelines model). A blogger who writes as titotal and describes themself as "a computational physicist" examined only the timelines forecast and its code, and wrote: "Not to mince words, I think it's pretty bad." The objections: "the fundamental structure of their model is highly questionable and at times barely justified, there is very little empirical validation of the model, and there are parts of the code that the write-up of the model straight up misrepresents."6 The biggest target was the "superexponential" assumption, under which each doubling of the length of tasks AI can do comes faster than the last. titotal also noted that "the scenario in the short story is not the median forecast for any AI futures author", and warned against "people taking shoddy toy models seriously and basing life decisions on them".6

"A sloth with an abacus" (Peter Johnson, as the authors summarised him). Before titotal, a Reddit and LessWrong commenter argued that the same superexponential trend is so strong that the model barely responds to its starting point: by the authors' own summary, "even starting at 'a sloth with an abacus' would give you AGI within a few years."8

"A bet against Amdahl's Law" (Steve Newman). Newman, whose own timelines "are substantially longer", read the takeoff forecast. He granted that AI 2027 "contains the best researched argument for short timelines", but argued that its huge speed-ups need every part of AI research to be automated at once: "Projecting speedups reaching 2000x is 'pricing to perfection'; if the ability to route around difficult-to-automate activities is anything short of perfect, progress will fall short of the anticipated speedup curves."7

What the authors replied

• -To Johnson (April 2025): "We admit that the superexponential trend was implemented in a quick-and-dirty way which doesn't let you apply it to arbitrary starting conditions, but we think it's a real possibility that accurately reflects recent trends." They added that a second method, "benchmark-and-gaps", was "the stronger informant of our views".8
• -To Newman (April 2025): they endorsed a reply by Ryan Greenblatt, who told Newman that "you mostly argue against large AI R&D multipliers, but you don't spend much time directly referencing compute bottlenecks".8
• -To titotal (December 2025): "we disagree with most of titotal's criticisms. While they pointed out a few mistakes, for which we are grateful, on the whole we think the problems that they pointed out do not add up to the model being 'bad.'" They paid titotal $500 "for pointing out some bugs and mistakes". Fixing the issues raised by titotal and Peter Johnson "pushed the (time horizon extension) SC median with Eli's parameters back from Aug 2027 to Nov 2029". (SC, "superhuman coder", is the model's key milestone.) On whether such models are worth making: "the alternative to modeling is vibes, and we think that vibes+modeling is better than just vibes alone."9 They also said they now think they should have given superexponential growth more weight, "even though our upcoming models' timelines are overall longer".9
• -On the AI 2027 page itself (22 Nov 2025): "we don't know exactly when AGI will be built. 2027 was our modal (most likely) year at the time of publication, our medians were somewhat longer."3

Where the dates went next

• -November 2025: Marcus wrote that the scenario "has officially been postponed", saying Daniel Kokotajlo "is now with me in thinking we probably won't see AGI this decade".10 The Kokotajlo post Marcus shows as a screenshot reads: "Things seem to be going somewhat slower than the AI 2027 scenario. Our timelines were longer than 2027 when we published and now they are a bit longer still; 'around 2030, lots of uncertainty though' is what I say these days."10
• -December 2025: the authors' new AI Futures Model gave "a roughly 3-5 year shift in our median for full coding automation" compared with the model behind AI 2027. They said this "has primarily come from improving our modeling of AI R&D automation".11
• -2026: the authors' timelines got shorter again. In August 2026, Kokotajlo's all-things-considered median for an Automated Coder was Nov 2027, and for superintelligence Mar 2029. Lifland's were Jan 2030 and Jul 2033.2 They now say "reality seems to be going about 70-90% as fast as AI 2027 predicted."2 The tracker's What its authors say section has the details.

How to read this. The critics and the authors agree on more than the headlines suggest. The authors say 2027 was only their most likely year, with medians later even then;3 in August 2026 only Kokotajlo's coder median (Nov 2027) still falls in 2027.2 The real disagreement is about method (can a model with few data points and a superexponential assumption tell you much?) and about bottlenecks (will automating AI research speed everything up, or will the parts that resist automation slow it down?). Watching the tracker's rows over the next year is the fairest test either side has offered.

When do the authors now expect superintelligence?

Not in 2027. In their August 2026 update, the authors' all-things-considered medians for superintelligence (ASI) are March 2029 for Kokotajlo and July 2033 for Lifland, as given in the post's charts.2 They describe these forecasts as "conditional on things going as fast as is technically feasible".2 These are the authors' dates for superhuman AI, not a forecast of an AI takeover.

How the scenario ends: the race and slowdown endings

AI 2027 branches into two endings. The authors wrote "the racing ending based on what seemed most plausible to us", then the slowdown ending "based on what we thought would most likely instead lead to an outcome where humans remain in control".3 In the race ending, a misaligned AI waits until it is trusted and then, in mid-2030, "finds the remaining humans too much of an impediment" and kills almost everyone. The authors add that "we do not endorse many of the choices made in either branch of this scenario".3 Both endings fall after 2027, so the tracker has no rows for them. For a full walkthrough, see AI 2027 explained.

Should you worry about AI 2027?

It is a scenario, and its authors say "AI 2027 is not a recommendation or exhortation. Our goal is predictive accuracy."3 They now think "reality seems to be going about 70-90% as fast as AI 2027 predicted", and their superintelligence medians run from 2029 to 2033.2 Every 2027 event in the tracker is still Not yet due. If the scenario is weighing on you, AI anxiety: what helps has practical steps; in the US you can call or text 988 (Suicide & Crisis Lifeline).