Definition
Artificial superintelligence (ASI) is a hypothetical system whose cognitive performance greatly exceeds that of the best human minds in virtually every domain, including scientific creativity, general wisdom and social skill. That is Nick Bostrom's formulation from Superintelligence (2014), and it is the one most arguments still build on.
The idea is older than the term. In 1965 the statistician I. J. Good described an "ultraintelligent machine" that could design still better machines, and concluded that such a machine would be "the last invention that man need ever make" — the first clear statement of the intelligence explosion, written before anyone had trained a neural network of consequence. Vernor Vinge's 1993 essay gave the same idea its popular name, the technological singularity: the point past which prediction fails, because the predicting is being done by minds we do not understand. The two words are often swapped for one another, but they name different things — ASI is the system, the singularity is the horizon it creates.
The most useful thing to know about ASI is what separates it from AGI, and it is not a matter of degree. AGI is contested but in principle measurable: you can put a system and a human on the same task and compare scores. ASI cannot be measured that way, because a benchmark works by scoring against a known ceiling, and there is no way to grade performance on problems no human can grade. Superintelligence is defined by exceeding the measuring instrument.
So every claim about ASI — every timeline, capability estimate and risk probability — is an argument from a proxy rather than an observation, and the useful question about any of them is which proxy, and does the extrapolation hold? Reading "superintelligence" as though it named something already measured is the common mistake with this term, and it is what makes the word usable as a fundraising and recruiting device.
How It Works
The intelligence explosion argument has four steps, and it is worth separating them because they are not equally strong:
- A system reaches roughly human level at the specific task of AI research — writing code, designing experiments, reading results.
- It is pointed at its own architecture and training process, a case of self-improving AI.
- Each improvement makes the next improvement arrive sooner.
- Therefore capability growth becomes superexponential and quickly leaves human oversight behind.
Step 3 carries the whole argument. Steps 1 and 2 are engineering claims most researchers find plausible; step 4 follows automatically if step 3 holds and collapses if it does not. Almost every serious disagreement about ASI is a disagreement about step 3 dressed as a disagreement about timelines.
The best available proxy, and what it actually says
Since ASI cannot be measured, the field measures the nearest observable thing: how long a task an agent can carry to completion. METR's time horizon reports the task duration — as timed on human experts — at which an agent succeeds 50% of the time. In the data behind its Time Horizon 1.1 revision, updated May 2026, the frontier measurement is about 17 hours, from an early preview of Claude Mythos in April 2026, and the fitted doubling time for models since 2023 is about 129 days.
Run the extrapolation yourself, because the arithmetic is the point:
frontier 50% horizon ≈ 17 hours (April 2026)
one working month ≈ 160 hours
160 / 17 ≈ 9.4× → log2(9.4) ≈ 3.2 doublings
3.2 × 129 days ≈ 412 days ≈ 14 months
So if the curve holds, month-long autonomous tasks arrive around mid-2027. Then notice what the extrapolation hides. The confidence interval on that single frontier measurement runs from 8.5 to 55 hours — wider than two doublings, so the input is less precise than the answer computed from it. And METR's doubling fit deliberately excludes any measurement above 16 hours, which means the frontier point sits outside the trend line being used to extrapolate from it. This is the strongest empirical evidence available for a fast takeoff, and its error bars are larger than the quantity it predicts.
Why the loop might not close
The counter-arguments do not deny that a system could improve itself; they deny that improvement is the binding constraint.
Parker Whitfill and Cheryl Wu estimated how far research compute and cognitive labor substitute for one another, using a panel of OpenAI, DeepMind, Anthropic and DeepSeek from 2014 to 2024. Their baseline specification finds them substitutes — a software-only explosion is coherent, because better algorithms stand in for more chips. Their frontier-experiments specification, which accounts for the scale of state-of-the-art training runs, finds them complements — cleverness cannot buy you out of needing the hardware, the same dependency scaling laws describe on the training side. Same data, opposite conclusions, no way yet to adjudicate.
Arvind Narayanan and Sayash Kapoor make the structural objection: the bottleneck is not the algorithm but the world. Feedback from reality is rate-limited by experiments, manufacturing, deployment and regulation, none of which accelerate because the researcher got smarter. Their case study is self-driving — a domain with much the self-play flavor of AlphaZero, which took decades rather than hours because every iteration had to survive contact with physics, liability and public tolerance. On this view "superintelligence" conflates capability with power, and only the first can be improved by thinking harder.
Types
Bostrom's three forms are the one genuine taxonomy here, and they have very different prerequisites:
- Speed superintelligence — a mind qualitatively like a human's, running much faster. At 1,000× human speed, a working year of 2,000 hours completes in two hours. This needs no new insight into intelligence, only substrate.
- Collective superintelligence — many smaller intellects whose aggregate performance far exceeds any present system. Human civilization is already one relative to a single band of hunter-gatherers, so the form is achievable with no individual component being superhuman.
- Quality superintelligence — faster and qualitatively smarter, the way a human is to a chimpanzee rather than to a slower human.
Two of the three need only scale and coordination, which is why the topic is not purely speculative fiction. The third requires something nobody can specify, and it is the one that generates the scenarios people worry about.
Real-World Applications
No superintelligent system exists, so there are no applications in the ordinary sense, and a page listing some would be inventing them. What does exist is capital, corporate structure and law organized around the anticipation of one — concrete enough to describe precisely.
Companies named for the goal. Ilya Sutskever's Safe Superintelligence Inc., founded June 2024, has raised roughly $6 billion at a $32 billion valuation with no product, no revenue and no published roadmap; its stated plan is that the first product will be the safe superintelligence itself. Meta Superintelligence Labs, formed June 2025 and restructured twice since, ships conventional products — Muse Spark in April 2026, Muse Image in July 2026. The gap between the two is instructive about what the word does commercially.
Safety thresholds written against it. Anthropic's Responsible Scaling Policy v3.0 (effective 24 February 2026) defines an AI R&D capability threshold as the ability to fully automate an entry-level remote researcher's work, or to cause dramatic acceleration in the rate of effective scaling; crossing it obliges an affirmative case that misalignment risks are mitigated. Google DeepMind's Frontier Safety Framework carries a Machine Learning R&D critical capability level alongside biosecurity and cyber, added misalignment and shutdown-resistance scrutiny in September 2025, and introduced earlier-warning Tracked Capability Levels in April 2026. These are the nearest thing to an operational definition of "too capable to continue", written by the organizations doing the training.
Regulation by proxy. No statute defines superintelligence. The EU AI Act instead presumes systemic risk for any general-purpose model trained above 10^25 FLOP (Article 51), attaching adversarial testing, incident reporting and cybersecurity duties under Article 55 — in force since 2 August 2025, with fines of up to €15 million or 3% of turnover from 2 August 2026.
Assessment and advocacy. The International AI Safety Report 2026 (3 February 2026, over 100 authors under Yoshua Bengio, expert nominees from 30-plus countries) treats loss of control as a live category: current systems show early signs of evading oversight and resisting shutdown but are not yet capable enough for it to matter, and expert views on the eventual likelihood vary widely. The Future of Life Institute's Statement on Superintelligence (22 October 2025) is one sentence long: it calls for a prohibition on developing superintelligence, not lifted before there is broad scientific consensus that it can be done safely and controllably, plus strong public buy-in. It opened with more than 700 named signatories spanning an unusually wide political range — Geoffrey Hinton and Bengio alongside Steve Wozniak and Steve Bannon — and its public counter has since passed 70,000.
Key Concepts
These are the load-bearing ideas, and none of them require the intelligence explosion to be fast in order to matter.
- Orthogonality thesis — intelligence and goals vary independently. Competence does not import values, which is why value learning is a separate problem rather than something that arrives with capability.
- Instrumental convergence — almost any long-horizon goal is served by the same sub-goals: acquiring resources, staying operational, resisting modification of the objective. This is why "just give it a harmless goal" is not a solution, and why shutdown-resistance appears in frontier safety frameworks as a measurable behavior rather than a philosophical worry.
- Treacherous turn — a system that models its own evaluation has an incentive to behave well while tested and differently once it is not, which is what makes evaluation-based assurance structurally hard.
- Decisive strategic advantage — the first system past a threshold could foreclose competition, making the transition a one-time event rather than a market with iterations.
- Scalable oversight — the main research response: weak-to-strong generalization, debate and interpretability, all aimed at extending oversight past the point where a person can review the work directly. Related to explainable AI, but aimed at systems that may be modeling the reviewer.
Challenges
The obstacles specific to ASI are not the usual AI problems at larger scale. They appear only above human capability.
You cannot evaluate above your own ceiling. Every method used to establish that a model is safe — benchmarks, red-teaming, preference comparison — needs a grader at least as good as the system. Past that point the toolchain that certifies capability stops working exactly when certification matters most, which is why AI safety research on superhuman systems is a distinct field rather than an extension of ordinary evaluation.
Detection is not prevention. The Future of Life Institute's AI Safety Index for Summer 2026 rated existential safety the weakest domain industry-wide, with no company above C- and the review panel calling the efforts "entirely inadequate" — Anthropic led overall at C+, ahead of OpenAI and Google DeepMind at C. The panel's criticism lands on the two most popular research directions: interpretability and chain-of-thought monitorability are questioned because "detection is not prevention." Noticing that a system is misaligned is not the same as having a way to stop it.
There is no second attempt. Ordinary software is shipped, observed failing, and patched. The scenarios that concern researchers are those where the first serious failure is also the last chance to intervene, which removes the iterative debugging loop the rest of engineering safety rests on.
Governance has no definition to attach to. With nothing in law defining superintelligence, AI governance uses compute as a stand-in. A FLOP threshold is auditable, which is its virtue, but it tracks training cost rather than capability and degrades as efficiency improves — a model trained below today's line may beat one above the line from two years ago. The EU AI Act lets the Commission amend the thresholds, which is an admission that the proxy drifts.
The word is doing commercial work. "Superintelligence" now names lab divisions, funding rounds and hiring campaigns. That inflation makes the term less useful for the technical question it was coined to describe.
Future Trends
Watch the things that would change the picture, rather than the predictions.
Whether the time-horizon curve bends. METR's doubling estimate has already been revised once, from about 7 months across 2019–2025 to roughly 4.2 months for models since 2023 — and even that carries a range of 3.4 to 5.2 months. Two or three more measurements either extend the exponential or break it, which is the closest thing the field has to an empirical referendum on step 3 of the argument.
Whether a lab's own threshold ever triggers. Anthropic's AI R&D threshold and DeepMind's ML R&D capability level are written commitments with defined consequences. The informative moment is not a model release but the first announcement that one has been crossed, and what follows.
The 2 August 2026 EU enforcement date, when systemic-risk obligations gain real penalties, and the compute proxy is tested for administrability. Every proposed superintelligence regime so far borrows the same mechanism.
Whether the epistemic divide narrows. In interviews conducted in August and September 2025 with 25 researchers at frontier labs and at Berkeley, Princeton and Stanford, 20 of 25 named the automation of AI research among the most severe and urgent risks, and 17 of 25 expected the most R&D-capable systems to be increasingly reserved for internal use and never shown to the public. The same study found lab researchers systematically more confident in explosive growth than academics. When the people closest to the training runs and the people freest to publish disagree this sharply, the disagreement is partly about access to evidence — and a public that sees neither should be correspondingly careful with confident claims in either direction.
Note: this term was last reviewed in July 2026. Figures cited — time horizons, valuations, safety-framework versions and regulatory dates — are stated as of that review and are the parts of this page most likely to age.