When could superintelligence happen?

Two questions, not one

“When could superintelligence happen?” is really two questions, and the answers may be far apart.

The first is when AGI, roughly human-level generality, might arrive. The second is how long after that superintelligence, a large margin above the best humans, might follow. Some researchers think the second follows the first almost automatically, through the self-improvement loop I. J. Good described in 1965; others think the step is slow, or not guaranteed at all. So a forecast for one is not a forecast for the other, and most of the confident dates in circulation are for the first.

What the forecasters say

The forecasts below are given with who made them, when, and (because this is where most confusion starts) exactly what they were forecasting. They are not endorsed here; the point is the spread.

Dario Amodei, Anthropic. In an essay published in October 2024, “Machines of Loving Grace”, Amodei wrote that what he calls “powerful AI”, his own term, defined in the essay as a system smarter than a Nobel Prize winner across most relevant fields, “could come as early as 2026”. He does not use the word AGI for this, and says so.

Sam Altman, OpenAI. In a blog post dated 6 January 2025, “Reflections”, Altman wrote: “We are now confident we know how to build AGI as we have traditionally understood it.” OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work”. The post gives no date for AGI itself.

Demis Hassabis, Google DeepMind. In a WIRED interview published in 2025, Hassabis said: “In the next five to 10 years, there’s maybe a 50 percent chance that we’ll have what we define as AGI.” Google DeepMind’s working definition of AGI, published on 2 April 2025, is “AI that’s at least as capable as humans at most cognitive tasks”. That is a forecast for the second rung, AGI, not the third.

Surveyed researchers. The largest survey of the field, by Katja Grace and colleagues, published on arXiv in January 2024, collected responses from 2,778 researchers who had published at leading machine-learning venues. Aggregated, they gave a 50 per cent chance by 2047 of “high-level machine intelligence”, which the survey defined as unaided machines being able to outperform humans at every possible task. That is a stricter target than most company definitions of AGI, and this page does not treat the two as the same thing. The survey also reported that this date had come forward by thirteen years compared with the same survey a year earlier.

The consensus report. The International AI Safety Report 2026, published on 3 February 2026 under the chairmanship of Yoshua Bengio with more than 100 contributing experts, gives a range rather than a date. Its scenario analysis says that by 2030 “AI progress could plausibly range from stagnation to rapid improvement to levels that exceed human cognitive performance”, that “the trajectory of AI progress through 2030 is uncertain, but current trends are consistent with continued improvement”, and that “experts disagree about whether these capabilities will generalise beyond mathematics and programming”.

Superintelligence specifically. Dated forecasts for the third rung are rare. Safe Superintelligence Inc., whose website describes “one goal and one product: a safe superintelligence”, has given no date. Amodei’s essay of 12 September 2026, “We Must Pace the Frontier”, talks in terms of near-term harms (a “swarm” capable of “taking over the entire internet with a persistent botnet” within “6–12 months”) rather than a date for a system beyond human ability across the board. And the most widely reported number of September 2026 was a risk estimate, not a timeline: Reuters reported that Evan Hubinger, Anthropic’s alignment lead, estimated a greater than 10 per cent chance of AI causing human extinction within a decade. That is a statement about the probability of catastrophe; it is not a forecast that superintelligence will exist by a given date, nor evidence that AGI leads quickly to it.

Why they disagree

Three reasons account for most of the spread.

They are forecasting different things. OpenAI’s charter defines AGI by economic value; Google DeepMind by “most cognitive tasks”; the Grace survey asked about every task; Amodei’s “powerful AI” is something else again. A 2026 date for one target and a 2047 date for another may not conflict at all.

The recent pace can be read two ways. Systems that could not reliably add up in 2022 now pass professional examinations and carry out multi-step tasks with tools. One reading extrapolates that curve; the other notes what is still missing (continuous learning, reliability over long horizons, knowing when you are wrong) and argues the curve may flatten before it gets there.

Nobody knows whether scaling continues. Much of the progress since 2019 has come from making models larger and training them on more data and compute. Whether that keeps working, and for how long, is an open empirical question, and the answer moves every forecast at once.

A note on the administration’s “Super Intelligence”

A UK reader is most likely to reach this page because of the news of September 2026, so one clarification is needed before going further. The US administration’s “Super Intelligence” is a new name for artificial intelligence in general, the systems in use today. It followed two Truth Social polls (19 and 21 September 2026) and President Trump’s UN General Assembly address of 22 September 2026, and the White House’s own press release of that day refers to artificial intelligence as “hereinafter officially called ‘Super Intelligence.’” It is not a forecast, a scientific assessment, or a claim that AGI or superintelligence in the research sense has arrived or is expected by any date; nothing in the administration’s statements asserts a change in what the technology can do. The rename timeline records each step, and distinguishes what was polled, what was said in a speech, what the White House has published, and what, if anything, has become an executive order or directive or actual usage across federal documents.

How to hold the question

A defensible position, given the evidence above: AGI within the next decade is a serious possibility that serious people rate anywhere from likely to unlikely, depending partly on how they define it; superintelligence is a further step whose timing and even inevitability are disputed; and anyone offering a single confident date, in either direction, is telling you about their assumptions rather than about the world. This page will be reviewed quarterly and the forecasts updated as they move.