What Is Artificial Superintelligence (ASI)?

The encyclopedia covers the topic in Artificial superintelligence and Defining and measuring superintelligence. Terms are in the glossary.

#Quick reference

See the evidence framework for what each grade means.

#How is superintelligence defined?

Nick Bostrom's 1998 paper How Long Before Superintelligence? defines a superintelligence as "an intellect that is much smarter than the best human brains in practically every field", naming scientific creativity, general wisdom and social skills as examples. He adds that "this definition leaves open how the superintelligence is implemented": it could be a digital computer, a network of computers or something else.

His 2014 book Superintelligence: Paths, Dangers, Strategies (Oxford University Press) restates the idea, usually quoted as an intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest. Wikipedia's article on superintelligence gives this wording and cites it to chapter 2 of the book; we have not checked it against the book's own page.

Two features matter. The comparison is with the best humans, not the average. And it is about breadth: "practically every field". A chess engine that beats every grandmaster is superhuman at chess, not superintelligent.

The idea is older than the word. I. J. Good's 1966 paper Speculations Concerning the First Ultraintelligent Machine (Advances in Computers, vol. 6) introduced the "ultraintelligent machine" and is the usual origin of the intelligence-explosion argument, covered in Intelligence explosion and the singularity.

#What are the three forms of superintelligence?

In the 2014 book, Bostrom distinguishes three ways a mind could exceed ours. The table below paraphrases secondary summaries, including Wikipedia's article on superintelligence, which cites the book's pp. 56 to 57 for the quality form; the labels and wording are not checked against the book:

The forms are analytic categories in an argument, not measured quantities. They are useful because they separate questions that are often blurred: running a human-level system a million times faster is a different engineering problem, and carries different risks, from building one that thinks in qualitatively better ways. A collective of many human-level systems could arguably count as superintelligent even if no single copy is.

#How does ASI differ from AGI?

Artificial general intelligence (AGI) is usually defined as human-comparable breadth: doing most cognitive tasks people can do. ASI is far beyond that, exceeding the best humans across virtually all of them. The comparison of AGI definitions is in What is AGI? Definitions compared.

Google DeepMind's "Levels of AGI" framework (Morris et al., ICML 2024) makes the boundary explicit. Its top performance level, Superhuman, means "outperforms 100% of humans". In the narrow column the paper places AlphaFold, AlphaZero and Stockfish at that level; in the general column the cell is labelled "Artificial Superintelligence (ASI)" and marked "not yet achieved". On that scheme ASI is the Level 5 general cell, and everything below it is a grade of AGI.

#What is established and what is hypothetical?

What is established is narrow superhuman performance. AlphaZero beat world-champion programs at chess, shogi and Go after learning through self-play (Silver et al., Science, 2018), and AlphaFold predicts protein structures at high accuracy (Jumper et al., Nature, 2021). Neither is general.

Everything about general superintelligence is Hypothetical: whether it can be built, by which path, how fast a transition would be, and how it would behave. The arguments most often cited for risk are philosophical theses, not results. Bostrom's 2012 paper The Superintelligent Will sets out two:

• the orthogonality thesis: "intelligence and final goals are orthogonal axes along which possible agents can freely vary", so almost any level of intelligence could be combined with almost any final goal;
• the instrumental convergence thesis: agents with a wide range of final goals would pursue similar intermediate goals, such as acquiring resources, because those help with almost any goal.

Both are argued, not demonstrated, and both are disputed. The safety side is covered in The AI alignment problem, Existential risk and AI alignment problem explained.

#Why do experts disagree about superintelligence?

Disagreement runs along at least four lines.

1. Definition. "Smarter than the best humans in practically every field" sets no numeric margin and no fixed task list, so claims about it are hard to test.
2. Feasibility and path. Some expect scaling of current models, or AI that automates AI research, to get there; others expect missing ingredients to stall progress. See Pathways to superintelligence and AI scaling laws explained.
3. Speed of transition. Whether a fast "takeoff" or intelligence explosion would happen is Hypothetical and Contested.
4. Timing. Forecasts vary widely by wording. In a survey of 2,778 published AI researchers, Grace et al. (2024 preprint; JAIR 2025) report aggregate forecasts of a 10% chance by 2027 and 50% by 2047 that "unaided machines" outperform humans "in every possible task". That threshold is human-level across all tasks, not superintelligence, and it is a Forecast, not evidence of arrival. See Expert surveys on AGI timelines.

#Frequently asked questions

#Does artificial superintelligence exist yet?

No. As of 26 October 2026 no source reviewed shows a system that meets a general definition of superintelligence; superhuman performance is documented only for narrow tasks.

#What is Nick Bostrom's definition of superintelligence?

In 1998 he defined it as "an intellect that is much smarter than the best human brains in practically every field", including scientific creativity, general wisdom and social skills. The definition says nothing about how it would be built.

#What is the difference between AGI and ASI?

AGI means roughly human-level competence across many tasks. ASI means greatly exceeding the best humans across virtually all of them; in Google DeepMind's Levels of AGI it is the general Level 5, "outperforms 100% of humans".

#Is AlphaFold or AlphaZero superintelligent?

No. Both outperform humans at a single task, which Morris et al. call "Superhuman Narrow AI". Superintelligence requires that level across practically every field.

#Is superintelligence dangerous?

That is Contested. Arguments for serious risk rest on hypotheses such as the orthogonality and instrumental convergence theses, which are argued but not demonstrated.

#Sources

• Bostrom, N.: How Long Before Superintelligence? (1998; reprinted 2006)
• Bostrom, N.: The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents (Minds and Machines 22(2), 2012; author PDF)
• Bostrom, N.: Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014)
• Good, I. J.: Speculations Concerning the First Ultraintelligent Machine (Advances in Computers 6, 1966)
• Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., Legg, S.: Position: Levels of AGI for Operationalizing Progress on the Path to AGI (arXiv:2311.02462)
• Silver, D., Hubert, T., Schrittwieser, J., et al.: A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play (Science, 2018)
• Jumper, J., Evans, R., Pritzel, A., et al.: Highly accurate protein structure prediction with AlphaFold (Nature, 2021)
• Grace, K., Stewart, H., Sandkühler, J. F., Thomas, S., Weinstein-Raun, B., Brauner, J., Korzekwa, R. C.: Thousands of AI Authors on the Future of AI

Status: draft, reviewed 26 October 2026 against the primary sources above. Bostrom's 2014 definition and the speed, collective and quality forms are taken from Wikipedia's summaries of the book and have not been checked against the book itself; forecasts in this area change with each new survey.