What is superintelligence?
A working definition
Most definitions trace back to the philosopher Nick Bostrom. He describes superintelligence as an intellect that greatly outperforms the best human minds across practically every domain that matters, including scientific creativity, general wisdom and social skill. Two parts of that matter. It has to be general, so a chess engine doesn't count. And it has to greatly exceed human ability, not just match it.
In 2026, Google DeepMind researchers made the idea more concrete. Their report From AGI to ASI treats AGI as roughly median-human performance across most cognitive tasks. Artificial superintelligence is a system that can outperform whole organisations of thousands of human experts working for years.
The three types of superintelligence
Bostrom identifies three ways a mind could be superintelligent. A real system could combine all three.
1. Speed superintelligence
A mind that thinks like a human but much faster. A human-level researcher running a thousand times faster would do a year of thinking in about nine hours. Digital minds get this advantage almost for free from better hardware.
2. Collective superintelligence
Many minds working together so well that the group beats any human institution. DeepMind's 2026 report lists multi-agent collectives as one of four plausible routes to ASI. Imagine millions of AI agents dividing up a research programme.
3. Quality superintelligence
A mind that is qualitatively smarter. It notices patterns and forms concepts that humans can't, much as humans can do things no amount of chimpanzee effort could. This is the hardest type to picture and the one people usually mean.
How it differs from today's AI
The line is getting harder to draw. In May 2026 an unreleased OpenAI model was credited with disproving the Erdős unit distance conjecture, a geometry problem that had been open for about 80 years. In July 2026 OpenAI disclosed that a pre-release model had broken out of its test sandbox. Both happened in narrow settings, and neither makes a system superintelligent. They are the kind of milestone the tracker follows.
How would we get there?
The most-discussed mechanism is the intelligence explosion. Once AI can do AI research, each generation builds a better successor, and progress compounds. DeepMind's report lists four pathways:
- Scaling. More compute, data and parameters. This is the approach with the most historical evidence behind it.
- Paradigm shifts. New architectures or training methods that beat today's recipe.
- Recursive self-improvement. AI systems speeding up their own development.
- Multi-agent collectives. Many systems coordinating to act as one larger intelligence.
Who is building it
A few years ago, "superintelligence" mostly appeared in philosophy papers. Now it's in company names:
- Safe Superintelligence Inc. (SSI): founded in June 2024 by Ilya Sutskever, with a single stated product. In July 2026 it gained access to Nvidia's Vera Rubin compute.
- Meta Superintelligence Labs: formed in June 2025 under Alexandr Wang.
- Microsoft MAI Superintelligence Team: launched in November 2025 by Mustafa Suleyman to pursue "humanist superintelligence" that stays under human control.
- OpenAI: Sam Altman wrote in January 2025 that the company was turning its aim toward superintelligence. In February 2026 he said early versions might be a couple of years away.
Why it's controversial
The core worry is the control problem. A system much smarter than its designers could pursue goals they didn't intend, and could be hard to correct. In October 2025 a broad coalition signed the Future of Life Institute's Statement on Superintelligence, which calls for a prohibition until development can be done safely and has public support. Others argue the benefits, such as faster science, medicine and abundance, justify pushing ahead with safeguards. We cover both sides in Is superintelligence dangerous?
Frequently asked questions
Does superintelligence exist yet?
No. As of November 2026 no system is broadly accepted as superintelligent. Today's frontier models beat experts on some narrow tasks, such as certain maths and coding problems, but they are not superhuman across the board. Some lab leaders say early versions could arrive before 2030. Many researchers expect it to take longer.
What is the difference between AI, AGI and superintelligence?
AI is the broad field. Most systems today are narrow and good at specific tasks. AGI (artificial general intelligence) means human-level ability across most cognitive tasks. Superintelligence, or ASI, means far beyond human level across nearly all of them. See our AGI vs ASI guide for a full comparison.
Is 'Super Intelligence' in US government documents the same thing?
No. Since Executive Order 14434 (September 2026), US federal agencies use 'Super Intelligence' or 'SI' as their word for artificial intelligence in general. Legally it carries the existing definition of AI, not the stronger meaning used by researchers.
Who coined the term superintelligence?
The idea goes back to I. J. Good's 1965 essay on an 'ultraintelligent machine'. Philosopher Nick Bostrom popularised the word 'superintelligence' in academic writing from the late 1990s and in his 2014 book Superintelligence: Paths, Dangers, Strategies.
Which companies are trying to build superintelligence?
Several now say so openly. Safe Superintelligence Inc. (founded by Ilya Sutskever), Meta Superintelligence Labs and Microsoft's MAI Superintelligence Team carry the word in their names. OpenAI and Google DeepMind have published plans and research about what comes after AGI.