A forecast called AI 2027 predicts that superintelligent machines arrive in 2027 and the world ends shortly afterwards

The title came loose from the document

Most of what this site covers begins with somebody inventing something. This one begins with a real document, written by people using their own names, published openly, and available to anyone who wants to read it. Almost nothing has been fabricated. The distortion happened in transit.

AI 2027 went up in April 2025. It is a scenario: month by month from 2025 forward, AI agents becoming useful, then AI systems doing the work of AI researchers, then an intelligence explosion late in 2027. It was written by Daniel Kokotajlo, who had resigned from OpenAI’s governance team the previous year, together with Eli Lifland, Thomas Larsen and Romeo Dean, with Scott Alexander editing.

Then it escaped its readership. It was discussed in policy circles and reportedly read by the US vice president, and somewhere in that journey the title stopped being the name of a document and became a date. What people search for now is the year: is AI 2027 on track, is it realistic, is it coming true, will the world end in 2027. A respectable share of the people typing that have never opened the thing they are asking about, which is not a criticism of them. It is how numbers travel.

It ends twice, and one ending is fine

The first thing the document does that the date cannot is branch.

In one ending, competition between states and labs drives capability past the point where anyone retains control, and it goes the way the frightening summaries say. In the other, coordinated action slows the race, oversight holds, and the catastrophe does not arrive. Both are written out. Neither is marked as the expected one.

The authors are also explicit about what kind of object this is. They describe the narratives as hypothetical, and the project as a planning tool rather than a literal forecast of events. That is an odd sentence to have to quote about a document that has frightened a lot of people, and it is the single most load-bearing fact on this page.

The clearest statement of it comes from a critic rather than from the authors. In April 2025 the Machine Intelligence Research Institute published a response written by someone who agrees that catastrophe is plausible on roughly this timescale, and who put his own most likely year at 2028. He wrote that, as far as he could tell, nobody associated with AI 2027 actually expected things to move as fast as depicted, and that the scenario was meant as a story about how things could plausibly go fast. When a sympathetic reader who shares your conclusion says your document is not a prediction, that is worth more than any number of hostile reviews.

2027 is the tallest bar, not the expected year

There is a second reason the number sticks, and it is a genuinely interesting one about how people read charts.

When AI 2027 was published, aggregated forecasting markets on transformative AI had a shape worth looking at. The peak of the distribution, the single most likely year, sat in or near the end of 2027. The midpoint of the same distribution, the year by which half the probability had accumulated, sat in 2031.

Those are four years apart, and both are correct. A distribution with a long tail to the right does that: the most likely single outcome arrives early, while the expectation sits much later, because the possibilities that things take much longer outweigh the possibilities that they take much less. The mode is a headline. The median is a forecast.

So a person who reads that the tallest bar is 2027 has not misread the chart. They have stopped at the first number. And the scenario’s title, by naming that year, hands them a reason to stop there.

One caveat the MIRI piece raises against itself, which is worth carrying: a superhuman coder and an artificial general intelligence are not the same thing, and comparing forecasts of the two is not like for like. The four-year gap is real. Its precise size depends on what you are measuring.

The revisions, which are the tell in the other direction

This site keeps a page on Baba Vanga, whose annual prediction lists are republished each year with the items unchanged and the dates advanced. Alien contact was promised for 2025, did not happen, and reappeared for 2026 with a month attached. A prophecy’s date only ever moves forward, and it moves quietly.

AI 2027 has been revised twice in public and the revisions went in opposite directions. In December 2025 the authors wrote that developments since publication suggested a slower path to autonomous AI research than they had assumed. In April 2026 they wrote that further developments in AI programming tools suggested a pace closer to the original scenario. They have said they intend to keep integrating new evidence.

Then, in July 2026, they published a follow-up called AI 2040: Plan A, arguing for slowing the research timeline, forcing transparency from the labs and spreading power across more companies and countries. The new number is a policy horizon. It is worth saying clearly, because the title invites the opposite reading: the apocalypse was not rescheduled to 2040.

And in April 2026 Kokotajlo went back to an essay he wrote in 2021 about what 2026 would look like, and marked it. Forecasts of this kind are almost never graded, least of all by their authors, and whatever the verdict, the willingness to publish one is the clearest available difference between this and a prophecy.

What can honestly be said before the year arrives

Not much, and a page that claimed otherwise would be committing the error it is describing.

The decisive claims in AI 2027 sit in late 2027. That has not happened. Anyone telling you now that the forecast has failed is reading the future, and anyone telling you it is confirmed is doing the same. The honest position is that the year is ahead and the question is open.

What can be established is the difference between the document and the date. It is a scenario, not a prediction, and its authors say so. It has two endings and one of them is fine. It names a year and not a day, so it will expire by fading rather than by failing. The number that made it famous is the peak of a distribution whose midpoint lies years later. And its authors keep publicly changing their minds in both directions, which is what distinguishes a forecast from a revelation.

None of that settles whether fast timelines are right. Serious people disagree about it, including people who think the danger is real and the schedule wrong. What it settles is narrower: the thing frightening people in a search box is not quite the thing that was published, and the gap between them is a year that somebody wrote on a cover.