In short
Superintelligence is a hypothetical intellect that greatly exceeds the best human minds across virtually all domains of interest, including scientific discovery, reasoning, creativity and planning. In AI the term for the machine form is artificial superintelligence (ASI). Today's systems beat humans at particular tasks, but no source reviewed here shows a system that meets the broad definition. The 2026 U.S. executive order on "Super Intelligence" changed government terminology; it did not show that ASI exists.
#Key takeaways
- Definition: an intellect that greatly exceeds human cognitive performance in virtually all domains of interest (Bostrom, 2014).
- AGI vs ASI: AGI means broad, roughly human-level ability. ASI means performance well beyond it. No agreed scientific threshold separates the two.
- Status: hypothetical. Narrow superhuman performance is documented; general superintelligence is not.
- Paths: researchers discuss scaling AGI, new AI paradigms, recursive improvement and multi-agent collectives. None is a proven route.
- Main challenge: keeping more capable systems safe, controllable and accountable.
#1. What superintelligence means
The modern discussion is closely tied to philosopher Nick Bostrom and his book Superintelligence: Paths, Dangers, Strategies (2014), which defines a superintelligence as "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest."
The key phrase is virtually all domains. A chess engine that beats every human is superhuman at chess, yet it is not superintelligent, because its advantage stops at the board. The definition also says nothing about consciousness or how the system is built: it is about capability.
A system would need more than high benchmark scores or fast processing to qualify. The properties usually listed are:
- Superior reasoning: solving problems beyond what human experts can.
- Scientific creativity: proposing original hypotheses, theories and experiments.
- Transferable learning: carrying knowledge into unfamiliar fields.
- Strategic judgment: anticipating consequences in complex systems.
- Technological invention: designing new software, materials and machines.
These are properties a superintelligence would be expected to have. They are not verified capabilities of any existing system. For a longer treatment of the definition and how to measure it, see Defining and Measuring Superintelligence.
#2. AI vs AGI vs ASI
| AI | AGI | ASI | |
|---|---|---|---|
| Meaning | The broad field of building systems that do tasks associated with intelligence | General ability to learn and reason across many tasks, at about human level | Ability well beyond the best humans across virtually all domains |
| Breadth | Often task-specific, increasingly general | Broad | Broad |
| Level | Varies; sometimes superhuman on one task | Roughly human | Far above human |
| Status | Deployed | Contested; no agreed test | Hypothetical |
The three describe different things. AI is the technology category. AGI is about breadth at roughly human level. ASI is about depth far beyond it. Researchers disagree on the benchmarks that would establish AGI; Google DeepMind's Levels of AGI paper (Morris et al., 2023) proposes grading both performance and generality instead of using a single line. The articles Artificial General Intelligence and Definitions and Operational Criteria for AGI cover the AGI side, and the post AI vs AGI vs SI vs ASI compares all four terms.
#3. Three forms of superintelligence
Bostrom distinguishes three ways a system could exceed humans:
- Speed superintelligence: human-style thinking run far faster, so a year of research fits in a short time.
- Collective superintelligence: many agents or processes coordinated so the whole outperforms any individual or current organization. This links to multi-agent AI research.
- Quality superintelligence: better thinking in kind, able to grasp or solve what the best humans cannot.
These are conceptual categories, not a classification of any current product.
#4. How could ASI emerge?
No scientific roadmap guarantees it. A June 2026 Google DeepMind report, From AGI to ASI (Genewein et al.; arXiv 2606.12683), describes four potential pathways from AGI to ASI: scaling AGI, AI paradigm shifts, recursive improvement, and ASI emerging from multi-agent collectives. It also covers likely obstacles and argues the change may arrive as a series of societal shifts rather than one step.
Scaling. Better architectures, training, compute, data, reasoning, memory and tool use could keep raising general capability. More scale alone does not guarantee superintelligence.
Recursive self-improvement. An AI system helps improve itself or build a better successor, which then does the same. If each round speeds the next, the loop is sometimes called an intelligence explosion. It is a hypothesis, not a result: hardware limits, research bottlenecks, diminishing returns and safety requirements could slow or stop it. See The Intelligence Explosion Hypothesis.
AI-assisted AI development. This is observable today and should not be confused with the loop above. Anthropic's essay When AI builds itself reports that, by its own account, Claude writes over 80% of the code merged into its codebase, and calls this a step toward recursive self-improvement. The same piece lists limits: the productivity figures rest partly on staff surveys, humans still choose the research problems and scoring, and it gives no timeline for full recursive self-improvement. AI helping with research is not an AI independently running the whole cycle.
Multi-agent systems. Specialized agents for research, coding, planning and verification, coordinated by another system, might together exceed any single model, if they become reliable enough.
For the wider map of routes, including brain emulation and biological enhancement, see Pathways to Superintelligence.
#5. What could superintelligence do?
If it were possible and built safely, the proposed uses cluster around a few areas:
| Domain | Possible uses |
|---|---|
| Science and mathematics | New hypotheses, proofs and connections between fields |
| Medicine and biotechnology | Disease mechanisms, drug and molecular design |
| Energy and climate | Better energy systems, climate and environmental modelling |
| Engineering | New materials, chips, robotics and manufacturing |
| Cybersecurity | Vulnerability discovery and defence |
| Economics and planning | Forecasting, logistics and resource allocation |
These are potential applications, not evidence that existing systems can do them at superhuman level. Predictions from any such system would still need experimental validation, clinical testing and regulatory review. Better analysis also does not guarantee good values or good decisions; those depend on objectives, constraints and governance.
#6. What are the risks?
- Alignment: a capable system might pursue an objective in a way its designers did not intend. The more capable and autonomous the system, the more this matters. See The AI Alignment Problem and the post AI Alignment Problem Explained.
- Loss of control: if a system plans and adapts faster than its overseers, it becomes harder to evaluate, intervene on or shut down. This is a research and governance challenge, not a certain outcome. See AI Alignment and Control.
- Concentration of power: whoever has access could gain large economic, political or military advantages.
- Misuse and cybersecurity: capable systems can raise the scale of attacks as well as the strength of defences.
- Economic disruption: knowledge work could change substantially; scale and timing are uncertain.
#7. Does superintelligence exist today?
No. No source reviewed for this article shows a system that exceeds the best human minds across virtually all domains. Current AI is strong in language, coding, mathematics and tool use, and sometimes beats humans on specific tasks, but individual results do not add up to general superiority.
It helps to separate three things: what deployed systems can demonstrably do, new abilities researchers see appearing, and hypothetical superintelligence that has not been shown.
#Why "Super Intelligence" is in the news
NASA's page What is Super Intelligence? (updated 9 October 2026) cites Executive Order 14434, Inaugurating the Era of Super Intelligence, dated 29 September 2026, and says NASA is updating its communications to use the term "super intelligence" as the order directs. On that page the term is the broad category of systems that handle complex tasks usually needing human reasoning, and NASA relies on the statutory AI definition in 15 U.S.C. § 9401(3).
So the federal usage is a terminology choice for ordinary AI. It is not the research definition above, and it is not a claim that ASI has been achieved. Political wording, commercial branding and scientific definitions should not be treated as interchangeable.
#8. When might ASI arrive?
No one knows. Estimates differ because researchers disagree about what AGI requires, how fast capability will improve, whether AI can run its own research, and what limits exist in compute, energy, data and regulation. Treat dates as scenarios, not milestones. The post Expert Surveys on AGI Timelines summarizes the published forecasts and their limits.
#Frequently asked questions
#What is superintelligence in simple terms?
A hypothetical form of intelligence far more capable than the best human minds across nearly every important intellectual field.
#What does ASI stand for?
Artificial superintelligence: a hypothetical AI system that greatly exceeds human intellectual ability across virtually all domains.
#What is the difference between AGI and ASI?
AGI means broad intellectual ability comparable to humans. ASI means ability that greatly surpasses it.
#Is ChatGPT superintelligent?
No. It is capable across many tasks, but nothing shows it exceeds the best humans across virtually all domains.
#Does superintelligence exist today?
No source reviewed shows that it does.
#What is recursive self-improvement?
A process in which an AI system helps improve itself or build a more capable successor, possibly creating a feedback loop. It is a hypothesis, and AI assisting with research today is not the same thing.
#What is an intelligence explosion?
A hypothetical period of rapidly compounding gains in machine intelligence, often linked to recursive self-improvement. It is not an established or inevitable outcome.
#What are the three types of superintelligence?
Bostrom's speed, collective and quality superintelligence.
#Does superintelligence require consciousness?
No. The standard definition concerns capability. Whether machines can be conscious is a separate question.
#Who introduced the modern concept?
I. J. Good described an "ultraintelligent machine" in the 1960s, and Bostrom's 2014 book brought the modern framing to a wide audience.
#Sources
- Nick Bostrom, Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014), and How Long Before Superintelligence? (1998).
- Google DeepMind, From AGI to ASI (12 June 2026; arXiv 2606.12683).
- Meredith Ringel Morris et al., Levels of AGI for Operationalizing Progress on the Path to AGI (2023).
- Anthropic, When AI builds itself (update dated 18 September 2026).
- NASA, What is Super Intelligence? (updated 9 October 2026).
Superintelligence and AGI are subjects of ongoing research and debate. The capabilities, pathways and scenarios above are hypothetical and are not evidence that ASI exists or that its arrival is inevitable. The Anthropic figure is the company's own report and was not independently checked.
