AI's 2027 Doomsday Date Is a Scenario, Not a Prophecy

Type '2027' into a search box on an AI news site and a particular dread comes back. That is the year, a certain corner of the internet insists, when artificial intelligence slips past human control. The number is oddly specific, and it did not come from nowhere. It comes from a single document, and reading what that document actually says is the fastest way to calm down.

The source is AI 2027, a 71-page report published in April 2025 by Daniel Kokotajlo, a former OpenAI researcher, together with Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean under the banner of the AI Futures Project. It lays out a month-by-month story in which AI agents grow capable through 2026, coding becomes largely automated in early 2027, and a rapid loop of AI improving AI produces something like superintelligence by the end of that year. From there the report branches into two endings, one where humans keep control and one where they very much do not.

What the authors actually claim

Here is the part the scary headlines tend to drop. The authors do not say this will happen. They call it a scenario, a concrete guess offered to make an abstract debate easier to argue with. The point of writing a specific timeline was to give people something to push against, not to set a date. Kokotajlo has put real probabilities behind fast timelines and a serious chance of a bad outcome, which is why the report carries weight. But a forecast built from a chain of 'and then' steps compounds its uncertainty at every link, and the authors know it. Notably, the AI Futures Project has since pushed parts of its own timeline later, toward the end of the decade. A prophecy does not get revised. A working estimate does.

The wider track record should temper things further. Predictions of when machines will match humans have been made confidently for seventy years, and they have been wrong in both directions, too eager and too cautious by turns. Real constraints sit in the way of any smooth ascent: the supply of high-quality training data, the cost and power draw of compute, and the stubborn gap between acing a benchmark and doing reliable work in the world. None of that rules out fast progress. It just means a tidy 2027 cliff-edge is one path among many, not the track we are locked onto.

So treat the number for what it is. The honest worry underneath AI 2027, that capability may be outrunning our ability to steer it, is worth taking seriously, and plenty of level-headed researchers do. We have looked before at how easily that concern tips into fatalism. A date on a calendar is not evidence. It is a prompt to think, and thinking is the opposite of panicking.