What is superintelligence
Superintelligence refers to a hypothetical intelligence that would greatly exceed the best human experts in practically all significant cognitive activities — scientific research, engineering design, strategic planning and learning new skills. In the context of artificial intelligence it is abbreviated ASI, for artificial superintelligence.
The term was popularised by the philosopher Nick Bostrom in his book Superintelligence (2014); his notion rests on superiority across cognitive domains, as his earlier paper also explains.
What matters is the combination of breadth and level. The calculator on your phone adds faster than the best mathematician. Your satnav knows the way better than a taxi driver. A chess app beats the world champion. Yet none of them is superintelligent — each does one thing. A calculator cannot help you settle an argument with your in-laws, and a chess engine cannot run a research institute.
A more useful picture than an all-knowing robot is a team of exceptional scientists, programmers and strategists that can share findings instantly and learn new fields. A superintelligence would not need a human shape, would not have to live in a single computer and would not know everything. It would still need information — and it could still be wrong.
Likewise, greater intelligence does not automatically mean consciousness, goodness or unlimited power. Being able to solve problems is different from experiencing anything. And a brilliant plan does not mean you have the machines, energy, money and permissions to carry it out.
Superintelligence does not exist today. There is no public evidence of a system meeting such a broad definition. So why talk about it now? Because AI is improving faster than almost anyone expected — and rules, institutions and technical safeguards develop more slowly than new software.
AI, AGI and ASI: three floors, not a fixed ladder
Artificial intelligence (AI) is an umbrella term — it covers specialised systems as well as models usable for many different tasks. It helps to picture three floors of one house:
A system for a limited domain: chess, face recognition, film recommendations, checking products on a production line. Superhuman performance in one domain guarantees nothing elsewhere.
The ability to learn and handle a wide range of tasks at roughly human level. Beware: there is no generally accepted threshold and no single decisive test.
Clear superiority over the best humans across practically all cognitive domains. A hypothetical category — not a synonym for a powerful chatbot.
Today's general-purpose models do not fit neatly into “narrow or general”, though. One model can translate, program and describe a photo — and still fail at a project that takes several days. Breadth, level of performance and autonomy each develop differently; they are assessed separately in, for example, the research proposal Levels of AGI. So the house has no fixed staircase: some rooms on the first floor are already lived in, others do not even have a floor yet.
Bostrom also distinguishes how a machine could be smarter:
These forms can combine. But it does not follow that every system climbs the three floors and inevitably reaches the roof. More: AI, AGI and ASI without shortcuts.
Topics
Pick what interests you. The pages follow on from each other, so you can also read them one after another.
- 01How AI worksGuessing the next word, billions of knobs, training and why models get things wrong.
- 02AI layers and reasoningFrom chip to agent — and how a model “thinks” before it answers.
- 03From chess to chatbotsMilestones from Turing to the Nobel Prizes, the famous move 37 and why a machine could win.
- 04AI, AGI and ASIWhat the abbreviations mean and why they are not a fixed ladder.
- 05When and how it might arrivePaths to superintelligence, researchers’ estimates and how to spot real progress.
- 06Sceptical voicesWhy some experts do not expect superintelligence soon — and where they agree with the rest.
- 07AI agentsWhen AI not only answers but acts — and why permissions matter.
- 08AlphaFoldHow AI solved a fifty-year-old problem in biology — and why a protein’s shape is not yet a drug.
- 09Benefits and risksScience and education — but also King Midas, paperclips and gorillas.
- 10AI safetyAlignment, interpretability, rules and practical advice.
- 11GlossaryTerms from AGI to prompt injection, each in one sentence.
When might it arrive?
Nobody has a reliable date. The largest survey of 2,778 AI researchers (collected 2023) put a 50% chance on machines doing every task better and more cheaply than humans by 2047 — 13 years earlier than the same survey a year before. But it is an aggregate of estimates, not a forecast, and a lot depends on how the question is asked.
How to read those numbers and tell real progress from hype: When and how it might arrive.
FAQ
Is ChatGPT, Claude or Gemini a superintelligence?
No. They are services and families of models with many versions, and there is no public evidence that any meets the broad definition of ASI. A superhuman result on a single test does not justify that conclusion. They are, however, a good illustration of how fast AI is moving — ten years ago most experts would have placed them in the distant future.Does a superintelligence have to be conscious?
No. The definition of ASI does not require consciousness — performance and subjective experience are different questions. A model saying “I feel something” is not proof in itself, any more than an excellent maths result is.Does AI understand what it says?
It depends what we mean by “understand”. A model can use concepts in new contexts — explain a joke or spot a flaw in an argument. But that does not directly imply a human kind of understanding or inner experience. In practice you can measure correctness, transfer of knowledge and the ability to fix mistakes.Couldn't we just switch a dangerous system off?
For a limited system, usually yes. It is harder when a system is extensively connected, critical services depend on it or it can be copied. Research on *corrigibility* studies how to preserve the ability to correct and intervene even with capable systems. That does not mean today's chatbots can prevent being switched off.Wouldn't Asimov's laws of robotics be enough?
Asimov's stories are really a collection of tales about how such laws fail. What exactly counts as “harm”? An unpleasant truth? Surgery? Stopping someone climbing a dangerous mountain? The terms need interpretation and can conflict — which is precisely the alignment problem. Reliable behaviour needs training, testing and limited permissions.What is the technological singularity?
A hypothetical turning point after which technological development would far exceed our ability to predict it — much as we cannot see beyond the horizon of a black hole. The term was popularised by the mathematician and writer Vernor Vinge in his essay [The Coming Technological Singularity](https://edoras.sdsu.edu/~vinge/misc/singularity.html) (1993). It is often linked to the intelligence explosion, but it is not the same as AGI nor a proven phenomenon.Will AI take people's jobs?
It can automate parts of jobs and whole positions, but the extent and pace depend on capabilities, cost, organisation and rules. New roles can appear as others disappear. An overall benefit to the economy does not guarantee a smooth transition for a particular person.Does it make sense to prepare if we don't know whether ASI will arrive?
Yes. Better security, honest testing, good education and clear responsibility are valuable even if progress is slower. Preparedness can grow with the evidence — it need not rest on one bold prediction.Should I be afraid?
Fear does not help much; interest does. The shape of one of the most important technologies of our time is decided by people — developers, companies, politicians and the public. The more people understand what is at stake, the better the decisions we can make together.Sources and further reading
Sources for specific claims are linked directly in the text. For continuous reading:
Books to start with
- Melanie Mitchell: Artificial Intelligence: A Guide for Thinking Humans. Farrar, Straus and Giroux, 2019 — what AI can and cannot do, written for non-specialists and with critical distance.
- Brian Christian: The Alignment Problem: Machine Learning and Human Values. W. W. Norton, 2020 — alignment told through real stories from research.
- Stuart Russell: Human Compatible: Artificial Intelligence and the Problem of Control. Viking, 2019 — an accessible account of the control problem by the author of the most widely used AI textbook.
- Arvind Narayanan, Sayash Kapoor: AI Snake Oil. Princeton University Press, 2024 — how to tell what AI really does from marketing.
- Nick Bostrom: Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014 — the book that popularised the concept of superintelligence; a more demanding read.
Research papers and overviews
- I. J. Good: Speculations Concerning the First Ultraintelligent Machine. Advances in Computers, 1965 — the first description of the intelligence explosion.
- K. Grace et al.: Thousands of AI Authors on the Future of AI, 2024 — a survey of opinions, not a verified forecast.
- A. Narayanan, S. Kapoor: AI as Normal Technology, 2025 — a counterweight to the idea of a sudden break.
- International AI Safety Report 2026 — an overview of evidence and uncertainties about general-purpose AI.
- NIST AI Risk Management Framework — a framework for managing risk in practical deployment.
In Czech
- Jan Romportl: Superinteligence podle Jana Romportla I: Scénář vzniku (“Superintelligence according to Jan Romportl I: how it could emerge”). Novinky.cz, 16 October 2018.
- Jan Romportl: Superinteligence podle Jana Romportla II: Scénář převzetí moci (“II: a scenario of taking power”). Novinky.cz, 23 October 2018.
- Jan Romportl: Obavy ze superinteligence jsou namístě (“Concerns about superintelligence are justified”). Interview, Science Café, 7 October 2018.
- NeurIPS Test of Time Award for Tomáš Mikolov and his team. CIIRC CTU, 25 January 2024.
AI research in the Czech Republic takes place at, for example, CIIRC CTU, the AI Center at CTU's Faculty of Electrical Engineering, the Institute of Formal and Applied Linguistics at Charles University (language technology for Czech) and the Prague initiative prg.ai.
The views of experts who do not consider superintelligence close are summarised on Sceptical voices.