Types of superintelligence

Nick Bostrom’s 2014 book sorts superintelligence into three forms: speed superintelligence, a human-like mind running much faster; collective superintelligence, many minds working as one system; and quality superintelligence, a mind that can do what humans cannot. Other schemes sort it by path, breadth or role. None of these types describes a system that exists as of October 2026.

Why sort superintelligence into types at all?

Superintelligence is sorted into types because “smarter than humans” can mean very different things, and the differences change the risks.

A mind that thinks a million times faster than a person, a network of thousands of ordinary minds, and a single mind that grasps ideas no human can are all “superintelligent” in some sense. They would not behave alike, and they would not be controlled the same way.

The classic definition, explained in the definition guide, says only that a superintelligence greatly exceeds humans in virtually all domains. The taxonomies below ask how.

What are Bostrom’s three forms of superintelligence?

Bostrom’s three forms are speed superintelligence, collective superintelligence and quality superintelligence, set out in chapter 3 of Superintelligence: Paths, Dangers, Strategies (2014). The short descriptions below follow the summary written by Katja Grace for a 2014 reading group on the book.

Speed superintelligence

A speed superintelligence could do what a human mind does, but much faster.

The simplest picture is an emulated human brain running on faster hardware. A mind running thousands of times faster than a person would experience the outside world as nearly frozen. A day of human time would give it years of subjective thought.

Bostrom’s 1998 paper used a related idea: a mind with about the same abilities as a human brain, but much faster, which he noted “has been called” a “weak superintelligence.”

Collective superintelligence

A collective superintelligence is made of many smaller intellects that, if they communicate and coordinate well enough, act together with far greater capability than any one of them.

It is strongest on work that can be divided into parts and done in parallel. It can grow by adding more members or by organizing them better.

This form raises a boundary question. Bostrom’s 1998 paper said companies and scientific communities are not superintelligences because they are not intellects. The collective form asks how far better organization would have to go before a group counts as one mind.

Quality superintelligence

A quality superintelligence can carry out intellectual tasks that humans cannot do in practice, without necessarily being better or faster at the tasks humans can already do.

Bostrom compares the gap to the one between humans and other animals. Speed alone does not let a dog do algebra. Quality is the form most people picture when they hear the word, and it is the hardest to reason about, because by definition it can do things humans cannot specify in advance.

Are the three forms really different?

In the long run, Bostrom argues, not very.

The reading-group summary describes his distinction between direct reach and indirect reach. Each form is best at different tasks, so their direct reach differs. But each could build the others: a fast mind could design a better one, and a large collective could do the same. Their indirect reach is therefore the same.

In Bostrom’s account, then, the three forms are different starting points that may lead to the same place.

Why would digital minds have an edge?

Bostrom argues that a digital mind would have two kinds of advantage over a biological one, in hardware and in software.

• Hardware. Computer components can run far faster than neurons and send signals far faster than nerves. A digital mind is not limited to what fits in a skull.
• Software. A digital mind can be copied, edited without harm and, in some designs, can share memories between copies. It can add specialized modules, for example for vision.

These advantages are one reason much of the writing on superintelligence focuses on machines rather than on enhanced humans.

What paths could lead to superintelligence?

The paths usually listed are machine intelligence, biological enhancement, brain–computer interfaces and networks of people and computers. Two early lists, Vinge’s and Bostrom’s, are close to each other.

Vinge’s four paths (1993)

In “The Coming Technological Singularity,” Vernor Vinge listed four ways superhuman intelligence could arise.

1. Computers that are “awake” and superhumanly intelligent.
2. Large computer networks, with their users, that “wake up” as a superhuman entity.
3. Computer–human interfaces so intimate that users count as superhumanly intelligent.
4. Biological science that improves the natural human intellect.

Bostrom’s paths (2014)

Bostrom’s book covers a similar set:

• Artificial intelligence built from the ground up.
• Whole brain emulation, in which a brain is scanned in detail and run as software. Anders Sandberg and Bostrom set out a technical roadmap for this in a 2008 Future of Humanity Institute report.
• Biological enhancement, such as embryo selection or genetic engineering. Bostrom doubts that drugs or gene therapy in adults would raise intelligence far, but he describes repeated embryo selection as capable of large gains over generations.
• Brain–computer interfaces.
• Networks and organizations, such as better-coordinated humans or the internet acting as a collective.

Most of today’s public debate concerns only the first path. When AI companies, statements and bills discuss superintelligence, they mean machine superintelligence. The older literature treats it as one route among several.

Can superintelligence be narrow?

Under the classic definition, no. Under some newer uses, yes.

Google DeepMind’s 2023 “Levels of AGI” paper separates performance from breadth. Its top performance level, Superhuman, means outperforming 100% of humans. A narrow system can reach it: the paper lists AlphaFold, AlphaZero and Stockfish as superhuman narrow AI. Only a general system at that level is called artificial superintelligence, and the paper lists it as not yet achieved.

Microsoft AI uses the word in a narrower way. Its November 2025 “humanist superintelligence” announcement names “medical superintelligence” as one of its application areas: deliberately bounded, domain-specific systems. The company definitions guide compares this with other labs.

Both uses are now common. A reader should check which one a speaker means.

What roles might a superintelligence play?

Bostrom sorts possible systems by what they are allowed to do, in chapter 10 of his book. He calls these roles “castes.”

• Oracle: a system that only answers questions.
• Genie: a system that carries out one high-level command, then waits for the next.
• Sovereign: a system that acts autonomously in the world in pursuit of potentially long-range objectives.
• Tool: a system that is not designed to act as an agent at all.

The castes matter for safety. An oracle seems safer than a sovereign because it does not act directly. Bostrom’s point, as summarized by the reading group, is that each caste has its own failure modes; an oracle can still be misused even when it works as intended. The concern is covered in the debate guide.

How fast could a superintelligence emerge?

Takeoff is the time between roughly human-level AI and superintelligence, and Bostrom’s chapter 4 describes three cases.

• Slow takeoff: decades or centuries. Society has time to respond.
• Moderate takeoff: months or years. There is some chance to respond, but little time to analyze.
• Fast takeoff: minutes to days. There is little or no time to respond once it starts.

The fast case is what writers mean by an “intelligence explosion” or a hard takeoff. The intelligence explosion guide covers the argument and the objections.

Are there other taxonomies?

Yes. Three others come up often.

• Chalmers’ levels. In 2010, philosopher David Chalmers labeled human-level AI as AI, greater than human level as AI+, and far greater as AI++. AI++ is his term for superintelligence.
• Who it serves. Companies now sort by purpose: Meta’s “personal superintelligence,” Microsoft’s “humanist superintelligence,” and Safe Superintelligence Inc.’s “safe superintelligence.” These are goals, not technical categories.
• Unipolar or multipolar. Bostrom also asks whether one system ends up dominant, a singleton, or many compete. He defined a singleton in a 2006 paper as a world order with a single decision-making agency at the highest level.