Is Superintelligence Really Arriving in 2027?

Ask around and you will hear it stated as fact: superintelligence is coming in 2027. The year has a strange gravity to it. It shows up in policy arguments, investor decks, and the anxious corner of the internet, usually with the confidence of a train timetable. It is worth slowing down on, because the date is a forecast, not a schedule, and the people who made it famous no longer stand behind it in the way the meme suggests.

Where the number came from

The specific year traces largely to AI 2027, a detailed scenario published by a group of researchers that walked through how superhuman coding, autonomous research agents, and eventually superintelligence might arrive across 2027 and 2028. It was careful, well argued, and explicitly a scenario rather than a prediction. Somewhere between publication and the group chat, the caveats fell away and the year stayed.

The chorus around it is real. The chief executives of OpenAI, Google DeepMind, and Anthropic have all said they expect systems as capable as a human across most tasks within about five years. When the people building the technology talk that way, a fixed date starts to feel less like speculation and more like an announcement.

The forecasters moved the date

Here is the part that rarely travels with the headline. The AI 2027 authors have already revised their own timelines. Daniel Kokotajlo, one of the scenario's authors, has said his personal median has slipped to around 2030, and his co-author Eli Lifland has described the group's median moving back by roughly three years, per a review by FutureSearch. The team has set out the revision in its own words. 2027 remains possible in their view. It is no longer the center of the estimate.

Broader forecasts sit further out still. Forecasters on the prediction platform Metaculus have put roughly a 25 percent chance on human-level AI by 2029 and about even odds by 2033, a dramatic pull-forward from a decade ago but a long way from next year.

And some doubt the whole premise

Then there are the researchers who think the timeline debate is aimed at the wrong target. Gary Marcus has argued for years that a 2027 leap was never coming and has warned against building national policy on a date that keeps sliding. Yann LeCun has made a deeper objection: that today's large language models, however impressive, are not on the road to general intelligence at all, and that a different architecture will be needed. If either is right, arguing about the year misses the point.

None of this proves superintelligence is far off. The honest position is uncertainty, and the range of credible estimates is genuinely wide. What does not survive contact with the evidence is the confident, load-bearing 2027. It is one plausible year among many, borrowed from a scenario its own authors have since revised. Treating it as settled is its own small myth, and a related one is the assumption that a bigger model is automatically a smarter one, or that a system fluent enough to sound human must therefore be reasoning the way a person does. Capability is harder to date, and harder to define, than a single year lets on.