Definitions of Superintelligence, Compared
Most definitions of superintelligence (SI) agree on two things: it would exceed the best human minds, and it would do so across nearly every intellectual domain, not just one. They disagree on how far beyond human level it has to be and on which tasks count.
This page puts the most-cited definitions side by side, from I. J. Good in 1965 to the 2023 Levels of AGI framework, so you can see exactly where they overlap and where they differ.
Six definitions side by side
| Source | Term used | Definition | Compared against | Scope |
|---|---|---|---|---|
| I. J. Good, 1965 | Ultraintelligent machine | “A machine that can far surpass all the intellectual activities of any man however clever.” | The cleverest individual person | All intellectual activities |
| Vernor Vinge, 1993 | Superhuman intelligence | Entities with greater-than-human intelligence, created by technology. Vinge predicted that “within thirty years, we will have the technological means to create superhuman intelligence.” | Humans in general | Not specified; an AI, augmented humans or networks could all qualify |
| Nick Bostrom, 1998 | Superintelligence | “An intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills.” | The best human brains | Practically every field, explicitly including social skills |
| Nick Bostrom, 2003 | Superintelligence | “Any intellect that vastly outperforms the best human brains in practically every field, including scientific creativity, general wisdom, and social skills.” | The best human brains | Practically every field |
| Nick Bostrom, 2014 | Superintelligence | “Any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest.” | Humans’ cognitive performance | Virtually all domains of interest |
| Morris and colleagues (Google DeepMind), 2023 | Artificial Superintelligence (ASI) | Level 5, “Superhuman”: a general system that “outperforms 100% of humans”, the highest combination of performance and generality in the framework. | Every human (100th percentile) | A wide range of general tasks; listed as “not yet achieved” |
What the definitions share
- Breadth, not a single skill. Every definition requires superiority across almost all intellectual work. A system that beats everyone at chess or protein folding is “superhuman narrow AI” in the Levels of AGI framework, not superintelligence.
- The benchmark is the best humans. Good’s “any man however clever”, Bostrom’s “best human brains” and the 100th-percentile threshold in Levels of AGI all set the bar at the top of human ability, not the average.
- The substrate is open. Most of these definitions do not require a computer. Bostrom notes that his definition leaves open how a superintelligence would be built, and Vinge considered augmented human intelligence as well. Artificial superintelligence (ASI) narrows the idea to machines.
Where they diverge
How far ahead. The wording ranges from “far surpass” (Good) and “much smarter” (Bostrom, 1998) to “vastly outperforms” (2003) and “greatly exceeds” (2014). Levels of AGI is the only definition with a measurable threshold: outperforming 100% of humans.
Which abilities count. Bostrom’s earlier definitions explicitly include social skills and general wisdom. His 2014 wording refers to “cognitive performance” in “domains of interest”, which leaves room to debate which domains matter.
What kind of superiority. In his 2014 book, Bostrom separates three forms: speed superintelligence (thinking like a human, but much faster), collective superintelligence (many smaller intellects working together) and quality superintelligence (thinking that is qualitatively better). A system could be superior in one form without the others.
Is it a threshold or a trajectory? Good and Vinge mostly described a turning point that could set off rapid further change (see intelligence explosion). Levels of AGI treats ASI as the top level of a graded scale, alongside AGI and narrower systems.
The working definition on this site
WHAT-IS.SI uses this definition: superintelligence is a hypothetical intelligence that would substantially exceed the best human cognitive abilities across virtually all relevant domains. It follows Bostrom’s 2014 wording on breadth, takes the best humans as the benchmark like Good and the 2003 definition do, and treats the question of whether such a system exists as open, as Levels of AGI does.
For the full explanation, read What Is Superintelligence? To see how SI differs from today’s systems, read AI vs Superintelligence.
How to cite this comparison
You are welcome to reuse the table, as long as you credit the original sources and link back to this page:
“Definitions of Superintelligence, Compared.” WHAT-IS.SI, edited by Michael Moritz, updated 7 October 2026. https://what-is.si/definitions-of-superintelligence
If you notice a missing definition or a misquotation, please use the contact details in the legal notice below.
Sources
These are the primary sources for the definitions above. Listing them does not imply that they endorse this site.
- I. J. Good, “Speculations Concerning the First Ultraintelligent Machine”, Advances in Computers, vol. 6 (1965). The original “ultraintelligent machine” argument.
- The Coming Technological Singularity (Vernor Vinge, 1993). The essay that made “the Singularity” a popular term.
- How Long Before Superintelligence? (Nick Bostrom, 1998). An early definition and forecast, presented as philosophical analysis.
- Ethical Issues in Advanced Artificial Intelligence (Nick Bostrom, 2003). A restated definition, followed by a discussion of its implications.
- Nick Bostrom, Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014), chapter 2. The definition quoted here plus the speed, collective and quality forms.
- Levels of AGI for Operationalizing Progress on the Path to AGI (Morris and colleagues, 2023). A matrix of performance and generality that places ASI at Level 5. It is a research framework, not a universally agreed standard.