What we actually mean by superintelligence
The word is used loosely, often as a synonym for "very good AI". A precise definition matters, because what we fear and what we hope for both depend on it.
“Superintelligence” is one of those words that seems obvious until you try to pin it down. It gets used for a chatbot that writes decent code, for a hypothetical machine god, and for everything in between. If this journal is going to spend its days on the subject, it should start by saying precisely what it means.
A working definition
The most widely cited definition comes from the philosopher Nick Bostrom, in his 2014 book Superintelligence: Paths, Dangers, Strategies: an intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest.
Three parts of that sentence do the heavy lifting.
- Greatly exceeds. Not “matches”, and not “slightly beats”. A system that is a little better than the best human is impressive, but it is not categorically different.
- Cognitive performance. The definition is about what a system can do (reason, plan, discover, persuade), not about whether it is conscious or has an inner life.
- Virtually all domains. A chess engine vastly outperforms every human at chess and is useless at everything else. Narrow superhuman skill is not superintelligence.
Three ways to be smarter
Bostrom also distinguishes three forms that superiority could take. They are worth keeping apart, because they fail and succeed in different ways.
Speed. A mind that thinks like a human, but much faster. If it could do in an hour what takes a person a year, it would out-produce any research team simply by having more subjective time.
Collective. Many minds working together so well that the whole far outperforms any individual. Human civilization is already a weak version of this; the question is how much tighter the coordination could become.
Quality. A mind that is not just faster or more numerous, but better: one that notices patterns and makes inferences that humans could not make at all, no matter how long they had.
Most public discussion blurs these together. It shouldn’t: a speed superintelligence is, at least in principle, something we can reason about by analogy with fast humans. A quality superintelligence is the one that could genuinely surprise us.
What the definition leaves out
A good definition is also clear about what it does not claim.
It does not say that superintelligence must be a single system. It could be a network of models, tools and people. It does not say it must have goals of its own, or want anything. And it says nothing about timelines. It only says what would count if and when it arrived.
A definition tells us what we are arguing about. It predicts nothing.
That last point matters. Many debates about “when superintelligence will arrive” are really debates about different definitions. Someone who counts impressive narrow performance will give an early date; someone who insists on broad, general superiority will give a late one. Both can be reasoning correctly from different starting points.
Why precision matters
It is tempting to treat all this as academic. It isn’t, for a simple reason: the risks and the benefits scale with generality.
A narrow system, however capable, can only surprise us within its narrow domain. A system that exceeds us across virtually all domains, including strategy, persuasion and science, is different in kind. It would be the first thing we have ever built that could out-think us about how to deal with it.
That is why the rest of this journal will keep returning to this definition. When we discuss alignment, forecasting, or governance, the question underneath is always the same: what happens when the most capable minds on the planet are no longer human ones?
