When someone says "artificial intelligence," they speak as if it were a single thing, one entity that will learn to play chess one day and decide the fate of humanity the next. But that singular word hides a mess. Under the same name live an astonishingly capable calculator, a promise engineers have chased for seventy years, and a science fiction creature that does not yet exist. Confusing them is no innocent slip: it is the root of nearly all the fear and nearly all the disappointment that surround this technology.
The types of artificial intelligence are sorted by one simple question: how broad is what the machine can do? From there come three names worth keeping straight (narrow AI, general AI and superintelligence) because they mark the difference between what is already in your pocket, what is being researched with real effort, and what remains pure speculation.
01 · what existsNarrow artificial intelligence: the specialist
Every piece of artificial intelligence you have ever used belongs to a single category: narrow AI, also called ANI (Artificial Narrow Intelligence). The name is honest. Each system does one thing, or a handful of related things, and outside that lane it knows nothing.
The autocomplete that guesses your next word, the filter that sets junk mail aside, the model that translates from Spanish to Japanese, the assistant that transcribes a voice note, the program that beat the world Go champion: all of them are narrow. That a system plays Go better than any human does not mean it knows what a board is, or that it can hail a taxi. Its brilliance is intense but its edges are sharply drawn.
A narrow AI can outperform any human at its task and, at the very same time, be incapable of any other.
Here lives an expensive misunderstanding. When a language model converses with you fluently, the fluency deceives: it looks like a mind that understands everything. But it is still narrow. Its task, predicting likely text, turned out to be so broad that it brushes against many subjects, and that is why it gives the impression of generality. It is a specialist that learned to talk about almost anything without ceasing to be a specialist.
02 · what is pursuedGeneral artificial intelligence: the goal
General AI (AGI, Artificial General Intelligence) is the name of an old ambition: a machine able to learn and reason about any intellectual task a human can take on, and to carry what it has learned from one domain to another without being retrained from scratch. Not one more specialist, but a universal learner.
The decisive difference from narrow AI is not power, but transfer. You learned to ride a bike and that helped you, vaguely, to skate. You read about the French Revolution and some of that helps you understand another revolution you never studied. An AGI would make that leap between domains fluidly. Today's systems, however brilliant in their lane, do not: a model trained to read X-rays does not wake up one day knowing tax law.
The aspiration is not new. In 1950, Alan Turing opened his celebrated paper with a question, "can machines think?", and, rather than define thinking, proposed an imitation game to sidestep the problem [1]. Ever since, the field has swung between the promise of the general and the practice of the narrow. Today the big labs declare AGI as an explicit goal, but it is worth reading those declarations calmly: there is no consensus on what would count as AGI, nor proof that today's methods are enough to get there.
There is a temptation to believe AGI is today's narrow AI but more powerful: more data, more parameters, more compute. Scale may help, and in fact it has brought surprises. But breadth is not the same as size. A system can be gigantic and still fail to transfer what it knows from one domain to another. That a road has taken us far does not guarantee it reaches the end.
03 · what does not existSuperintelligence: the hypothesis
Beyond AGI, some authors place superintelligence (ASI, Artificial Superintelligence): an intellect that would surpass the best human brain in practically everything, from science to persuasion. The philosopher Nick Bostrom defined it as that which "greatly exceeds the cognitive performance of humans in virtually all domains of interest" [2].
It is important to be clear about its status: superintelligence does not exist. It is not a delayed product or a beta version; it is a hypothesis. Much of the debate on existential risk rests on the idea that, once AGI is reached, the machine could improve itself in an accelerating spiral. It is a serious argument that deserves attention, but it is a chain of conditionals, not a description of something real. Treating it as imminent mistakes the map for the territory.
Confusing today's AI with the superintelligence of fiction is fearing a lion while you pet a very clever cat.
| Type | Scope | Status today | Examples |
|---|---|---|---|
| Narrow (ANI) | One task or a bounded domain; does not transfer outside it | Real and everywhere | Spam filter, translator, voice assistant, today's language models |
| General (AGI) | Any human intellectual task; transfers across domains | Research goal, not yet achieved | None confirmed |
| Superintelligence (ASI) | Surpasses the best human at practically everything | Hypothetical | None; a speculative concept |
04 · why it matters to youThe distinction that saves you a scare
All of this might look like a taxonomists' quarrel were it not for the fact that the confusion has very concrete effects. Whoever believes their voice assistant is a step toward superintelligence lives with a misdirected fear. Whoever believes a language model already "understands" like a person entrusts it with decisions it should not have, and is startled when it fails with poise at something elementary.
Knowing which box a tool falls into gives you the right measure. Faced with a narrow system (and today all of them are), the correct question is not "is it dangerously intelligent?", but "at what concrete task is it reliable and where do its edges begin?". That is a question you can actually answer, and one that yields sensible decisions: where to lean on it, where to verify, where not to delegate.
Artificial intelligence is not a single thing advancing toward one destination. It is one end of the spectrum, heavily populated and very useful; a center researched with effort and no guarantees; and a far end that for now lives only in books. Telling the three apart does not make you an expert, but it does give you back something scarce on this subject: proportion. And with proportion you think better (and fear less) than with awe.
Sources
- Turing, A. M. (1950). Computing Machinery and Intelligence. Mind, vol. LIX, no. 236, pp. 433-460. The paper that posed the question "can machines think?" and proposed the imitation game (today known as the Turing test).
- Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press. Source of the definition of superintelligence as an intellect that greatly exceeds human cognitive performance in virtually all domains.
- Russell, S. & Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th ed.). Pearson. The field's reference text; it distinguishes between special-purpose AI and the aspiration to general intelligence, and places the debate in historical context.