Superintelligence
Superintelligence is a hypothetical intelligence that greatly exceeds the best human cognitive performance across practically every important field, rather than merely outperforming people on one task. The concept is implementation-neutral: it could refer to one artificial system or another form of intellect and does not by definition require consciousness. No current benchmark, model label, or isolated superhuman result is an agreed test for superintelligence.
Origin and context
Bostrom's 1998 paper defined superintelligence and considered routes from human-level artificial intelligence to much more capable systems. His 2014 book brought the concept, possible development paths, control problems, and societal consequences into wider research and policy discussion. A separate group of researchers later analyzed a formalized containment problem through computability theory, demonstrating independent scholarly uptake. These works establish a durable concept, but their conditional arguments and forecasts are not evidence that such a system exists.
Why it matters
The term identifies a capability regime in which assumptions designed for ordinary software or even human-level systems may no longer hold. If a system could outperform expert humans across science, strategy, persuasion, and engineering, its speed, replication, and ability to discover new methods could change both benefits and risks. The concept therefore shapes work on alignment, control, access, monitoring, and international governance. Clear usage matters because calling every strong model superintelligent collapses a conditional long-range problem into current product marketing.
Example
A chess engine that defeats every human player is superhuman at chess but not superintelligent under the broad definition. A language model that scores above many people on several exams also does not qualify without evidence across the relevant range of cognitive fields, operating conditions, and novel tasks. A defensible claim would need an explicit capability scope, strong and independent evaluations, comparison with the best human performance, reliability evidence, and tests resistant to contamination and selective reporting.
How it differs
AGI
AGI generally emphasizes breadth and generality around a stated competence threshold. Superintelligence adds a much stronger performance condition: capability far beyond the best humans across very broad domains. An AGI, under some definitions, could exist without being superintelligent.
AI Takeoff Speed
Takeoff describes the speed and dynamics by which an AI system might improve from one capability regime to another. Superintelligence describes the hypothesized level reached. A fast or slow transition is a separate claim from whether the destination is possible.
Maturity and evidence
Maturity is rated 4 for the term, not for the technology. It has a stable core definition, multi-decade use, an influential academic book, and independent peer-reviewed analysis. Measurement thresholds, development paths, timelines, and control conclusions remain contested. The object is hypothetical, so evidence supports established discourse and research adoption rather than demonstrated realization.
Limits and open questions
Intelligence is multidimensional, and phrases such as practically every field still require choices about domains, tools, time, embodiment, social context, and reliability. The concept does not itself predict when or how superintelligence would emerge, whether it would be agentic, or what goals it would pursue. Formal results about a specified containment problem should not be generalized to every possible architecture or safeguard. Claims should separate definitions, empirical capability evidence, conditional arguments, and forecasts.
Related terms
References
- How Long Before Superintelligence?International Journal of Futures Studies / Nick Bostrom · 1998 · class A
- Superintelligence: Paths, Dangers, StrategiesOxford University Press / Nick Bostrom · 2014-07-03 · class A
- Superintelligence Cannot Be Contained: Lessons from Computability TheoryJournal of Artificial Intelligence Research · 2021-01-05 · class A
Last updated: 2026-09-07
This term is also covered in the Skills Atlas as ai risk management skill.
This term is also covered in the Skills Atlas as model evaluation skill.
This term is also covered in the Skills Atlas as ai ethics skill.