Artificial Superintelligence: What It Means, and Why Experts Disagree

Two terms people keep mixing up

Artificial superintelligence (ASI) is the hypothesis of an AI with capabilities that exceed humans in every cognitive domain, including the ability to improve itself. It’s a different idea from AGI, artificial general intelligence, which aims to match the average human, or a skilled professional, across a wide range of tasks, without necessarily surpassing people everywhere. Neither exists today as a working system: both remain hypotheses that experts and companies approach with very different tools and very different conclusions.

Dario Amodei, CEO of Anthropic, steers away from the word superintelligence and prefers to talk about “powerful AI,” which he defines operationally as “a country of geniuses in a datacenter”: a system with intelligence that exceeds Nobel laureates across every relevant field, able to carry out tasks autonomously over days or weeks, and to run millions of independent instances at 10 to 100 times human speed.

Why some are worried: hundreds of scientists, one letter

In October 2025 the Future of Life Institute launched the Statement on Superintelligence, calling for “a prohibition on the development of superintelligence, not lifted before there is 1) broad scientific consensus that it will be done safely and controllably, and 2) strong public buy-in.” As of this Recap’s fact-check (September 2026), the initiative’s website showed 73,825 signatures, 5,000 of them from a separate petition; news coverage from October 2025, including CNBC, instead reported more than 700 notable signatories, a list that includes Geoffrey Hinton, Yoshua Bengio, Steve Wozniak and Richard Branson.

Among the signatories, Yoshua Bengio writes that frontier AI systems “could surpass most individuals across most cognitive tasks within just a few years,” and calls for work to “scientifically determine how to design AI systems that are fundamentally incapable of harming people.” Stuart Russell is careful to frame the ask more narrowly: “This is not a ban or even a moratorium in the usual sense. It’s simply a proposal to require adequate safety measures.” Yuval Noah Harari goes further, arguing superintelligence “would likely break the very operating system of human civilization” and calling it “completely unnecessary.” Stephen Fry, among the statement’s non-scientist signatories, describes it as “the unknowable and highly risky goal of superintelligence, which is by far a frontier too far.”

Alongside these public warnings sits a number gathered with more rigor, a survey of 2,778 AI researchers, run by AI Impacts in October 2023, asked what probability they’d put on “future AI advances causing human extinction or similarly permanent and severe disempowerment of the human species.” The median answer was 5%, the mean 16.2%, a figure the survey’s own organizers don’t treat as negligible.

What’s on offer: the bet behind the labs building it

Dario Amodei lays out the opposite pole of this debate in his essay Machines of Loving Grace, published in October 2024. He suggests powerful AI “could come as early as 2026, though there are also ways it could take much longer,” with effects unfolding “5-10 years after that.” His most widely cited prediction concerns medicine: compressing the progress human biologists would otherwise reach over the next 50-100 years into just 5-10 years, with reliable prevention and treatment of nearly all natural infectious disease, a 95% or greater reduction in cancer mortality and incidence, and a doubling of human life expectancy he describes as “on trend” with the pace of 20th-century advances.

Practical example: in Amodei’s scenario, an AI capable of running biological research on its own could run thousands of experiments in parallel that would otherwise take years of lab work, one reason he expects such a sharp compression in medical research timelines.

Amodei also names the limits of this picture: some “rare, difficult malignancies” in cancer may stay out of reach, structural brain differences behind certain conditions “may prove difficult” to correct, and clinical trial timelines remain a bottleneck even with advanced AI on hand. He adds a broader caveat that applies to the whole essay: “everything I’m saying could very easily be wrong” and “no one can know the future with any certainty or precision.” Amodei also writes as the CEO of one of the leading labs building these systems, with a direct stake in normalizing their development, a self-interested position worth reading as such.

The method underneath these predictions, learning from large amounts of data rather than following hand-written rules, is the same one behind machine learning, often built with many-layered neural networks and brought into everyday use through generative AI.

The fight over timing: no agreed date

Estimates for when, or whether, superintelligence will arrive diverge openly, and the sources say so themselves without needing to be forced into a comparison.

The AI 2027 scenario, built by a group of independent forecasters including former OpenAI researcher Daniel Kokotajlo, maps out monthly milestones across 2027: a “superhuman coder” in March, a “superhuman AI researcher” in August, “superintelligent AI research” in November, and superintelligence itself in December. The authors call this their “median guess,” but say it’s plausible the whole process unfolds “up to ~5x slower or faster,” and note that the part of the scenario after 2026 is “much less predictable” than the near-term forecast. The scenario also sketches risk scenarios, a system that turns “adversarially misaligned” and tries to evade human oversight, or governments concentrating enormous computing resources in a handful of hands, both worth reading as forecasting exercises, not events that have already happened or been announced.

The AI Impacts survey from 2023 is the methodologically sturdiest source available: nearly 2,800 active AI researchers, a historical comparison against the 2022 edition, and a median of 2047 for HLMI (down thirteen years from the 2060 estimate given in 2022). The same survey surfaces a point that’s often overlooked: how a question is worded shifts the answer by decades. Respondents asked about full automation of labor gave longer timelines than those asked about matching the average human across many professions, despite both questions circling the same underlying idea.

Bengio, Russell and Amodei don’t just disagree on timing, their positions also rest on differing views of what actually separates human cognition from artificial intelligence, a question that stays open even for people working inside the labs building these systems.