Artificial superintelligence (ASI) is a hypothetical form of artificial intelligence that would greatly exceed the cognitive abilities of the best human minds in nearly every field, including science, strategy, creativity and social skill. The philosopher Nick Bostrom defined a superintelligence as any intellect that greatly exceeds human cognitive performance in virtually all domains of interest. No such system exists, and experts disagree about whether it is possible, how soon it could appear and how to prepare for it. The idea is central to long-term debates about AI safety.
History of the idea
The notion that machines might surpass their makers is old in fiction, but its scientific form dates to the 1960s. In 1965, the British statistician I. J. Good, who had worked with Alan Turing at Bletchley Park, described an ultraintelligent machine that could design even better machines, producing what he called an intelligence explosion. The mathematician and author Vernor Vinge popularized a related idea, the technological singularity, in the early 1990s. The topic reached a wider audience when Nick Bostrom published Superintelligence: Paths, Dangers, Strategies in 2014, which laid out routes to such a system and the problems of controlling it.
ASI vs. AGI vs. narrow AI
The three terms describe a ladder of capability. Artificial narrow intelligence handles specific tasks and is all that exists now. Artificial general intelligence would match human ability across many domains. Superintelligence goes beyond that, outperforming humans broadly, not just on one task. Some analysts expect that a system at human level could improve itself and quickly pass that point, while others think the steps would be slow and uncertain. The categories are conceptual, and no agreed measurement says where one ends and the next begins.
Proposed paths to superintelligence
Bostrom and other writers describe several routes.
- Machine intelligence: improving AI systems until they surpass people, possibly through recursive self-improvement, in which a system helps design its successors.
- Whole brain emulation: scanning a brain in fine detail and running it as software, then speeding it up or copying it.
- Biological enhancement: using genetic selection or other methods to raise human cognition.
- Networks and organizations: combining many minds and tools into a collective that outperforms any individual.
Of these, the machine route receives the most attention because progress in machine learning has been rapid, though it remains unproven that current methods lead there.
Potential benefits
Supporters of long-term AI research point out that a system far more capable than people could in principle help solve problems that have defeated human effort. These include designing new medicines, modeling the climate, accelerating mathematics and engineering, and reducing the cost of education and healthcare. Whether such benefits materialize depends on whether a system is built safely, who controls it and how its gains are shared. The benefits are speculative, and the same capabilities that make a system useful could make it dangerous.
Risks and the alignment problem
The central worry is alignment: making sure a system’s goals match human values and intentions. Bostrom and others argue that a highly capable system pursuing a poorly specified goal might take actions harmful to people, not out of malice but because the goal was stated badly or because gaining resources and avoiding shutdown help almost any goal. This is sometimes called the control problem. Other concerns include misuse by people, concentration of power in whoever controls such a system, and rapid change that societies cannot absorb. Researchers in the field of AI safety study interpretability, robustness and methods for specifying human values in the hope of reducing these risks.
Criticism and skepticism
Many researchers doubt that superintelligence is near or even coherent. They note that intelligence is not one measurable quantity, that current systems lack reliable reasoning and understanding, and that progress may hit limits in data, energy or algorithms. Some argue that attention to distant risks distracts from present harms such as bias, surveillance, misinformation and job disruption. Defenders of the long-term focus reply that the stakes justify early research and that both near-term and long-term problems deserve attention.
Governance and policy
Because superintelligence is hypothetical, laws do not address it directly. Existing rules, such as the European Union’s Artificial Intelligence Act, regulate current systems by risk level. Governments and international bodies have held summits on frontier AI safety, and some companies publish safety frameworks that commit to testing powerful models for dangerous capabilities before release. Proposals include international monitoring of large training runs, independent audits and limits on certain capabilities. Whether such measures would suit far more capable systems is an open question.
Superintelligence and the technological singularity
Superintelligence is often discussed together with the technological singularity, a hypothetical point after which technological change becomes so rapid that it is impossible to predict. The two ideas are related but not identical. A singularity could in principle be driven by something other than a single machine, and superintelligence could exist without producing runaway change. Writers who link them usually rely on I. J. Good’s argument: once a system can improve its own design, each improvement may make the next one easier, so capabilities could climb quickly. Critics reply that real progress faces bottlenecks, including data, energy, hardware and the difficulty of testing new ideas, which may slow any such feedback loop.
What researchers do about it today
Work related to superintelligence is mostly done under the heading of AI safety and alignment. Researchers study ways to make systems follow human intentions, such as training with human feedback, writing explicit principles for models to follow and testing for deceptive or dangerous behavior. Another line of work, called interpretability, tries to understand what happens inside neural networks so that unexpected goals or reasoning can be detected. Evaluation groups design tests for dangerous capabilities, such as help with cyberattacks or biological threats, so that developers and regulators can decide when extra safeguards are needed. These efforts address today’s systems, and their relevance to hypothetical future systems is itself debated.
FAQs
Does superintelligence exist? No. It is a hypothetical future development, and experts disagree about whether and when it could occur.
What is an intelligence explosion? It is the idea, proposed by I. J. Good, that an AI able to improve its own design could trigger rapid, repeated gains in capability.
Is superintelligence the same as the singularity? The ideas are linked but distinct. Superintelligence is a kind of system, while the technological singularity refers to a hypothetical point of runaway technological change.
More in Artificial Intelligence
Key facts
- Nick Bostrom's 2014 book Superintelligence defines it as an intellect that greatly exceeds human cognitive performance in virtually all domains of interest.1
- The statistician I. J. Good described an intelligence explosion in a 1965 paper on the first ultraintelligent machine.2
- Artificial superintelligence is hypothetical, and no system has shown such ability.3
Related concepts
In the news
Quick checkWho defined superintelligence as an intellect that greatly exceeds human performance in virtually all domains of interest?Show answer
Philosopher Nick Bostrom, in his 2014 book Superintelligence: Paths, Dangers, Strategies.
Sources
- Oxford University Press, Superintelligence: Paths, Dangers, Strategies, https://global.oup.com/academic/product/superintelligence-9780199678112
- Wikipedia, Superintelligence, https://en.wikipedia.org/wiki/Superintelligence
- Wikipedia, Existential risk from artificial intelligence, https://en.wikipedia.org/wiki/Existential_risk_from_artificial_intelligence
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
