Superintelligence is a hypothetical form of intelligence that far exceeds the best human minds across almost every area of thinking. It does not exist today. Here is what the term means, how it differs from AGI and current AI, what experts forecast, and answers to the questions people ask most.
What Does Superintelligence Mean?
Superintelligence refers to an intellect that outperforms the best human brains in practically every field, including scientific creativity, general wisdom, and social skills. Most discussion focuses on artificial systems, but the definition does not require a computer.
Where the Definition Comes From
The term was popularized by Oxford philosopher Nick Bostrom, who defines superintelligence as any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest. His 2014 book Superintelligence: Paths, Dangers, Strategies brought the idea into mainstream AI policy and safety debates.
Two features of this definition matter:
Breadth is required.
A system that beats humans at one task does not qualify. Bostrom has noted that IBM's chess computer Deep Blue is not a superintelligence, because it is only strong within one narrow domain, and even there it is not vastly superior to the best players.
The physical basis is open.
Bostrom's framing allows for a single digital computer, a network of computers, biological tissue, or something else entirely.
Bostrom also describes collective superintelligence: a large enough group of separate reasoning systems that coordinate well could together exceed the capability of any single member.
Superintelligence vs. AGI vs. Today's AI
These terms are often confused. They describe different levels of capability.
A simple way to remember the difference: AGI is a generalist that can match a capable person, while superintelligence is a generalist that outperforms the best people across nearly everything.
Some researchers expect AGI to lead to superintelligence. One argument, often called recursive self-improvement, holds that a sufficiently capable AI could help design better AI, which could then design better AI still. Supporters say this could compress development into a short period. Critics question whether such a feedback loop would work that way in practice.
Is ChatGPT a Superintelligence?
No. ChatGPT and other large language models fall under the narrow category in standard classifications. They can write, summarize, and reason about many topics, which makes them feel general, but they still have clear limits in reliability, memory, and independent goal pursuit. Calling a current model "super AI" overstates what it does.
The Types of AI: Why You See "7 Types" Online
Articles that list "7 types of AI" combine two separate classification schemes.
Capability-based types (ANI, AGI, ASI) are the scheme most often used in technology writing.
Functional types describe how a system works rather than how capable it is:
Reactive machines
have no memory and only respond to present inputs. Chess engines of the Deep Blue era are a classic example.
Limited memory
systems can use recent or historical data to inform decisions. Most present-day applications, including chatbots, virtual assistants, and self-driving vehicles, fall in this category.
Theory of mind
would involve understanding the thoughts, emotions, and intentions of others. This remains theoretical.
Self-aware AI
would have awareness of its own internal states. This also remains theoretical and is closely tied to unresolved questions about consciousness.
These are classification conventions, not established laws, and sources arrange them differently.
How Long Until Superintelligence?
No one knows. The honest answer is a range with wide uncertainty. Two kinds of evidence are most often cited.
Expert Surveys
AI Impacts, a research group, has surveyed machine learning researchers over several years. Its 2023 survey of 2,778 researchers who had published in top-tier AI venues produced an aggregate forecast of a 50% chance of high-level machine intelligence (HLMI) by 2047. That is roughly 13 years earlier than the 2022 survey's estimate.
The survey defines HLMI as the point when unaided machines can accomplish every task better and more cheaply than human workers. That is a different target from superintelligence. HLMI would likely come first, but the survey does not measure how long the gap between them would be. Forecasts also vary widely, and the way a question is framed changes the answer.
Scenario Forecasts
The AI 2027 report, published in April 2025 by the AI Futures Project, is the most discussed recent example. It is a month-by-month scenario, not a statistical forecast. Its lead author, Daniel Kokotajlo, is a former OpenAI researcher. The report describes 2027 as the most likely single year in the authors' forecast, not a median.
Later commentary gave different medians. Secondary sources report estimates for full superintelligence ranging from the early 2030s to the mid-2030s, and a 2029 median for AGI. Because these sources disagree, treat AI 2027 as one informed scenario among many, and check the authors' current statements for updated dates.
Bottom line: There is no scientific consensus on when superintelligence will arrive. Credible forecasts range from the late 2020s to later in the century, and some researchers doubt it will happen at all.
Is Superintelligent AI Possible?
This is an open question. Bostrom argues that artificial superintelligence is plausible and could arise through deliberate design or by accident. Others argue that general intelligence may be harder to build than narrow intelligence, or that it may not be enough to produce something far beyond human capability on its own. Some researchers also note that advanced systems could behave unpredictably.
Current evidence does not settle the question. Anyone who states that superintelligence is certain or impossible is going beyond what is known.
Why Superintelligence Matters for Safety
Much serious discussion of superintelligence concerns risk and governance rather than capability. Bostrom has argued that superintelligence may be the last invention humans ever need to make, which implies that a poorly aligned system could be very hard to correct afterward.
Safety researchers study how to specify goals reliably, how to keep powerful systems under meaningful human control, and how developers can coordinate. The concern is not that a superintelligence would necessarily be hostile. It is that it could pursue a goal in ways its designers did not intend.
The term also appears in corporate branding. In 2025, Meta created Meta Superintelligence Labs, an AI division. The use of the word in a product or division name does not indicate that any system has reached superintelligence.
Common Mistakes When Using the Term
Calling any strong AI "superintelligence."
A system that excels at one domain is narrow, however impressive.
Treating forecasts as settled dates.
Surveys and scenarios describe uncertainty, not schedules.
Assigning a single IQ score.
IQ tests are normed against human populations, so they cannot meaningfully describe a system that exceeds humans across domains.
Frequently Asked Questions
What is superintelligence?
Superintelligence is a hypothetical intellect that greatly exceeds human cognitive performance across nearly all domains, including scientific creativity, general wisdom, and social skills.
Is superintelligent AI possible?
It is an open question. There is no scientific consensus. Some researchers, including Nick Bostrom, consider it plausible, while others doubt that current approaches will lead there.
How long until superintelligence?
No one knows. Expert surveys and scenario forecasts range from the late 2020s to later in the century, with wide uncertainty. Treat any single date with caution.
Will AI be superintelligent by 2027?
There is no established evidence that it will. The 2027 date comes from the AI 2027 scenario, which its authors present as a what-if rather than a firm prediction.
What is the difference between AGI and superintelligence?
AGI refers to a system that matches human ability across many tasks. Superintelligence refers to a system that greatly exceeds the best human performance across nearly all of them. AGI is generally considered a possible step toward superintelligence, though the link is debated.
Is ChatGPT a super AI?
No. ChatGPT is generally classified as narrow AI. It handles many kinds of language tasks, but it has clear limits in reliability and independent goal pursuit, and it does not exceed human ability across all domains.
What IQ is superintelligence?
There is no meaningful answer. IQ tests are designed to compare people with other people, so they cannot describe a system that would exceed human performance across most domains. Any specific number is a rhetorical device rather than a measurement.
What are the types of AI?
Two common schemes are used. By capability, AI is divided into artificial narrow intelligence, artificial general intelligence, and artificial superintelligence. By function, it is divided into reactive machines, limited memory, theory of mind, and self-aware AI. Some articles combine these into a "7 types" list.
What is the 30% rule for AI?
It is not an established standard. The phrase appears in business articles, and authors define it differently. Some use it as a guideline to automate roughly 30% of routine tasks while keeping human oversight of the rest. Check the author's definition before relying on it.