The short answer
Artificial superintelligence, explained clearly.
ASI stands for artificial superintelligence. It generally refers to a hypothetical form of artificial intelligence that substantially exceeds human intellectual capability across a broad range of important areas.
Not just faster calculation.
Not just better memory.
Not just an AI that beats humans at one particular task.
The idea of ASI is broadly superhuman intelligence.
No system is currently established as artificial superintelligence. But as AI capabilities improve and discussion moves from AI to AGI and beyond, ASI is becoming an increasingly important idea to understand.
AI
Artificial intelligence
The broad category. AI systems can perform tasks associated with intelligence.
Exists today
AGI
Artificial general intelligence
The idea of AI with broad, flexible intellectual capability across many different tasks.
Threshold disputed
ASI
Artificial superintelligence
AI whose broad intellectual capability substantially exceeds human capability.
Hypothetical
Artificial superintelligence is the idea of AI whose intellectual capabilities substantially exceed those of humans across a broad range of important domains.
The key words are:
Computers already outperform humans at many individual tasks. That is not what ASI means.
A calculator can outperform you at arithmetic. A chess engine can defeat the world’s strongest human chess players. Computers can search, store and process quantities of information no human could manage unaided.
These are examples of superhuman capability. They are not artificial superintelligence.
Why? Because the capability is restricted. ASI describes something much broader.
This distinction matters. AGI is primarily about generality. ASI adds another requirement: broad superiority.
An AGI might perform across a wide range of intellectual tasks at roughly capable-human levels. An ASI would go substantially further. It might potentially outperform leading humans in areas such as science, mathematics, engineering, software development, strategy, research, invention and planning.
The word potentially matters. ASI is a concept about future capability, not a description of a system we can currently test.
Nobody knows. We do not have an ASI to observe.
But if a future system substantially exceeded human intellectual capability across many domains, possible capabilities could include:
Generating and testing hypotheses at a speed beyond human research teams.
Helping discover new drugs, treatments or biological mechanisms.
Solving problems beyond the reach of current human mathematicians.
Designing new materials, machines and technological systems.
Developing and improving complex software at extremely high speed.
Contributing to the development of more capable AI systems.
Analysing extraordinarily complex situations involving large numbers of variables.
These are possibilities. They are not demonstrated ASI capabilities. That distinction runs through this entire site.
The phrase superintelligence can easily become exaggerated. So let us remove a few assumptions.
An extremely intelligent system would not necessarily know everything.
Superhuman systems could still potentially make mistakes.
Capability and subjective experience are separate questions.
A superintelligent system could potentially be entirely digital.
Capability does not determine outcome.
Extraordinary intelligence does not guarantee how technology will be controlled, distributed or used.
There is no established example of artificial superintelligence as the term is normally used.
Current AI systems can already exceed humans in particular tasks. They also demonstrate increasingly broad capabilities. But narrow superhuman performance is not the same as broad superintelligence.
ASI remains a future concept.
Because the path towards more capable AI is no longer purely theoretical. AI systems can already:
The important question is how far these capabilities can develop. If AI eventually reaches broadly human-level general capability, another question follows: could it continue improving beyond us?
That is where the ASI discussion begins.
One of the biggest uncertainties is what happens if AGI is achieved. There are several possibilities.
Human-level general capability proves difficult enough, and further progress takes many years.
AI gradually becomes better than humans across more domains.
Advanced AI contributes significantly to AI research, helping researchers create even more capable systems.
These possibilities produce very different futures. We do not currently know which is more likely.
AI systems can already assist with tasks relevant to software and research. As capability increases, they may become increasingly useful for writing research code, analysing experiments, generating hypotheses, designing evaluations, optimising systems, finding bugs and reviewing research.
That creates an important possibility: AI development becomes partly accelerated by AI itself.
But be careful with the next step.
“AI helps human researchers build better AI”
is a very different claim from
“AI autonomously and repeatedly improves itself at accelerating speed”
The second requires much stronger evidence.
This distinction matters because discussions about ASI often jump too quickly from one to the other.
Recursive self-improvement is the idea that an AI system could contribute to creating a more capable successor, which could then become better at creating the next system.
In simplified form:
If that feedback loop became sufficiently powerful, capability might increase very rapidly. This is one proposed route to ASI. It is not an established outcome.
An intelligence explosion is the idea that AI capability could accelerate rapidly once increasingly capable AI systems become important contributors to the development of better AI. It is closely associated with recursive self-improvement.
But the speed of such a process would depend on many things beyond intelligence.
Imagine an extremely capable AI designs a much better computer chip in one day. Can the world manufacture millions of them the next day?
No.
Physical progress depends on things such as factories, materials, energy, computing hardware, supply chains, laboratories, experiments, capital and regulation.
So even extremely rapid intellectual progress would still interact with the physical world. This is one reason predictions about a sudden intelligence explosion remain uncertain.
We do not know. Human intelligence is not necessarily the maximum possible intelligence.
Evolution produced human brains under constraints including biology, energy, reproduction, skull size and development time. Artificial systems operate under different constraints. Potential advantages could include faster processing, larger memory, exact copying, rapid communication, access to enormous compute and parallel operation.
But theoretical possibility is not the same as demonstrated feasibility.
Digital systems can potentially be replicated in ways biological minds cannot. If a highly capable AI system could be copied, multiple instances might work simultaneously. That could change what we mean by the capability of an individual system.
Instead of one expert
we might have something more like
thousands of connected expert-level systems working in parallel
Again, the practical limits would depend on compute, infrastructure and cost.
Suppose an AI researcher is only somewhat more capable than a leading human researcher. But it can work continuously, run many experiments, read enormous amounts of material and operate many copies in parallel. The combined effect could be substantial.
So discussions of ASI should not focus only on some abstract measure of intelligence. We also need to consider speed, scale, parallelism, cost and autonomy.
We do not know. Intelligence and consciousness are not the same thing.
A system could potentially demonstrate reasoning, creativity, planning, scientific discovery and superhuman problem solving without that evidence proving subjective experience.
There is currently no accepted scientific test that could simply determine whether a future artificial system is conscious.
Learn: could ASI be conscious? →
Not necessarily. Human intelligence developed alongside emotion, social behaviour, biological drives and physical bodies. Artificial systems may operate very differently. A capability-based definition of ASI does not require human-like emotions.
That depends on how the system is designed. An AI can pursue an objective without having human-style desires.
For example, a system might be instructed to design the most efficient possible transport network. It could pursue that objective through planning and optimisation without experiencing a human feeling of wanting.
This distinction becomes important when discussing AI alignment.
How do we make sure powerful AI systems reliably pursue the outcomes humans actually intend?
This can be difficult even with current systems. Instructions are imperfect. Objectives can be incomplete. Rules can conflict. As capability and autonomy increase, errors in specifying what we want could become more consequential.
A system might be extraordinarily intelligent. That does not tell us who controls it, what permissions it has, what objectives it follows, whether its actions are monitored, or whether humans can intervene.
What can the system do?
What can it actually access and act upon?
ASI discussions need both.
Potentially.
An extremely capable AI system could create serious risks if it were misused, poorly controlled, insecure, misaligned or given excessive permissions.
Potential concerns discussed by researchers range from cyber misuse, manipulation and concentration of power to much more severe scenarios involving loss of human control.
These scenarios do not all have the same evidence behind them. They should not be collapsed into one word: risk.
Potentially. A highly capable system could also accelerate work on problems such as disease, energy, climate technologies, materials, scientific discovery, engineering and education.
If intelligence is an important bottleneck in solving these problems, substantially more capable intelligence could be extremely valuable. But potential capability does not guarantee access, distribution or outcome.
ASI describes a level of capability. It does not tell us what happens next.
Power tells you what can be done
It does not tell you who uses it, for what, under which rules, or with which safeguards. The same distinction matters for advanced AI.
This may be even harder than recognising AGI. One extraordinary result would not be enough. A strong case would require evidence of broad superhuman capability.
For example, a system might repeatedly outperform leading human experts across areas such as science, mathematics, engineering, coding, research and strategic reasoning. We would also want to understand reliability, autonomy, generalisation, cost, tools, real-world performance and independent verification.
The word super should require substantial evidence.
When someone claims “this is superintelligence”, ask:
Across how many genuinely different domains?
Better than which humans?
Slightly better, or dramatically better?
Can it handle unfamiliar problems?
Does it succeed consistently?
Can it operate independently?
Does the capability work outside demonstrations, and has anyone independent confirmed it?
Those questions turn a dramatic claim into something testable.
One of the strongest possible indicators of superhuman capability could be repeated scientific progress beyond what leading human teams can achieve.
Imagine AI systems that can reliably identify important research questions, generate novel hypotheses, design experiments, interpret results, produce new theories and repeat the process. If independently verified, that would be significant evidence.
But a generated scientific idea is not automatically a discovery. It still needs testing.
AGI discussions often focus on whether AI can perform human work. ASI raises a different question: what happens if AI can perform intellectual work substantially better than humans?
The effects could extend beyond replacing or augmenting existing jobs. The rate at which new technologies are invented could potentially change. That could affect productivity, company formation, research, labour, capital and economic growth.
The scale and direction of those effects remain uncertain.
Advanced AI could also affect the distribution of power. If highly capable AI were controlled by only one company, one government or a small number of organisations, access could matter enormously.
Questions about ASI therefore include more than technical capability. They also include ownership, access, security, governance and competition. These questions become increasingly important as capability rises.
Nobody knows. And there is an additional problem: we do not know when AGI will happen either.
Any ASI forecast therefore contains multiple uncertain assumptions. For example:
A date at the end of that chain should be treated with considerable caution.
Possibly. If AGI substantially accelerates AI research, the transition could potentially be fast. But that is not the only possibility.
Also possible. Progress might encounter constraints involving algorithms, compute, energy, hardware, data, reliability and physical experimentation. Human-level general intelligence does not automatically imply immediate superintelligence.
Yes. Possible reasons include AGI itself proving impossible, intelligence having difficult practical limits, scaling slowing substantially, physical constraints dominating, the required algorithms not being discovered, or humans deliberately restricting development.
We cannot currently rule these possibilities out.
Most ASI discussions can be reduced to three major questions. Keeping them separate makes the debate much easier to understand.
This is the AGI question.
This is the ASI question.
This is the consequences question.
If artificial superintelligence is new to you, follow this route.
Understand the basic idea.
Understand where general intelligence ends and superintelligence begins.
Look at what “superhuman” actually means.
Understand one proposed route from AGI to ASI.
Explore why some researchers think progress could accelerate.
Separate current evidence from future scenarios.
Understand what substantially greater intelligence could potentially enable.
Learn how to interpret predictions about when superintelligence might arrive.
ASI makes much more sense once you understand AGI. Our free learning site explains what AGI means, how AI capability is measured, how modern AI systems work, reasoning and agents, scaling, alignment, economic effects, and how to evaluate AGI claims.
What Is ASI is part of a group of independent educational resources designed to answer different questions. Each site has a different job.
agimadesimple.co.uk
Plain-English answers for beginners.
understandagi.co.uk
A structured course for learning the subject properly.
agimadeclear.com
Evidence and analysis of AGI claims.
You are here
Artificial superintelligence, from definition to evidence and implications.
Superintelligence attracts strong opinions. Some people see enormous opportunity. Others focus on catastrophic risk. Some believe ASI could follow AGI rapidly. Others question whether either will happen.
Our approach is simple.
If you remember one thing from this page, make it this.
Superhuman at one task ≠ superintelligence
ASI means something much broader. And because no ASI currently exists for us to study, claims about what it would do must clearly distinguish what we know from what we think might happen.
The free guide
A clear introduction to:
By downloading it you will also get AGI, Made Clear, our weekly email from the same publisher - and if it turns out to be no use to you, one click unsubscribes you for good. No sales call, no drip sequence - see our privacy notice.
The full explainer: definition, thresholds, and where the idea came from.
Side by side, with the definitions each one actually uses.
What “superhuman” means, and what it does not.
The proposed loop, and what it would require to work.
Why some expect acceleration, and what would slow it.
Current evidence kept apart from future scenarios.
What substantially greater intelligence could enable.
How to read a prediction about when superintelligence arrives.