Superintelligence: Why Might “When?” Be the Weakest Question?

What is striking about the superintelligence debate is that certainty in it exceeds evidence. We speak of it as an appointment on a calendar, when the truth is that no one holds that calendar.

What Is Meant by Superintelligence?

Not a machine faster than us at arithmetic; we have had those for decades. In the literature, superintelligence is a hypothetical mind that outperforms human minds across the board. The best-known theoretical treatment is Bostrom's book (Oxford University Press, 3 September 2014), which asks what happens if machines surpass humans in general intelligence. The publisher's description suggests that such a mind might become "beyond our control", and that humanity's fate would then depend on it as the gorillas' fate now depends on humans. The book's value, in my view (an assessment, not a quotation), is that it insisted the ethical and safety question precede the engineering one, whether or not the scenario ever comes to pass.

Dates That Move

The AI Impacts survey of 2022 put the midpoint of the probability of "high-level machine intelligence" at 2059, 37 years out. Another survey, run in 2023 and reported in a paper published in 2024, covered 2,778 researchers who had published at leading AI venues. It put the midpoint for machines outperforming humans at every task at 2047, with a 10% chance by 2027, and for full automation of all occupations at 2116.

The two surveys differ in wording and sample, so they should not be equated literally. But the direction and size of the difference, twelve years in a year, is enough to counsel humility in judgement. Between 38% and 51% of respondents gave at least a 10% chance to outcomes as bad as human extinction, and they disagreed over whether faster progress would be better for humanity or slower. The only consensus is that there is none.

A Statement in One Sentence

In May 2023 the Center for AI Safety published a sentence signed by researchers and company leaders, among them Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, Sam Altman and Dario Amodei: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Note that the sentence offers no probability and no date; it is a call to order priorities, not a forecast.

The Near Danger Does Not Wait for the Date

The International AI Safety Report (second edition, 3 February 2026), led by Yoshua Bengio and written by more than a hundred experts with the backing of more than thirty countries and organisations, says that "significant gaps remain" in defensive safeguards, that sophisticated attackers can often bypass current defences, and that the real-world effectiveness of many safeguards is uncertain. These are present concerns that do not depend on a superhuman mind appearing.

At the same time, Stanford's 2025 index records that models leap forward on many tests yet stumble on others such as PlanBench, and sometimes fail at logical tasks that have a verifiable correct answer. This reminds us that superiority does not run along a single ladder; the phrase "surpassing humans" always needs a question: at which task, and by what measure?

What to Do

• Ask about harms that already exist, such as bias, surveillance and misuse, and do not let the distant distract you from the near.
• Take dates as attributed estimates, and be wary of any categorical claim about an appointment.
• Put safety and governance first now, rather than waiting for the accident; remedy after the event costs more.
• Build public awareness that tells sound research from sensational headlines.

None knows the unseen but God, and we are bidden to take the means: to understand what is in our hands and to build its safeguards before the need for them intensifies. The nearest sense of what is right is that responsibility precedes capability.