Types of AI: Narrow, General, Super (and Where We Actually Are)

The standard taxonomy of AI has three categories arranged on a capability ladder: narrow AI, general AI, superintelligent AI. The textbook taxonomy has four, replacing capability with cognitive style: reactive machines, limited-memory systems, theory-of-mind systems, self-aware systems. Both schemes are used in the field, neither maps cleanly to what current systems actually do, and both end up confusing as much as they clarify. The honest 2026 answer is that the categories are useful as a vocabulary for talking about what we have built and what we have not, but the boundary between them is fuzzy and the systems we ship do not sit neatly inside one box. This guide walks through both taxonomies, places real 2026 systems against them, and explains where the labels start to break down.

Table of contents

• The standard taxonomy
• Narrow AI examples in 2026
• Why "general" is harder to define than people think
• Reactive, limited-memory, theory-of-mind, self-aware
• The 2026 state -- where current models sit
• Other useful taxonomies
• Frequently asked questions
• The bottom line

The standard taxonomy

The most-cited modern taxonomy, popularised by writers like Nick Bostrom and used widely in tech press, has three rungs:

• Artificial Narrow Intelligence (ANI), sometimes called "weak AI": systems that perform a specific task at human level or above, but only that task
• Artificial General Intelligence (AGI), sometimes "strong AI": systems that can learn and perform any intellectual task a human can
• Artificial Superintelligence (ASI): systems that exceed humans at essentially all cognitive tasks, including the task of building better systems

This taxonomy is useful for discussing capability scope. It has two big problems. First, it is binary inside each rung -- you are either AGI or you are not -- when actual capability is continuous and unevenly distributed across tasks. Second, the boundary between narrow and general is exactly where the current debate is, and the taxonomy provides no internal vocabulary for that boundary. A model that handles thousands of tasks competently is technically "narrow" if you read the definition strictly, which is clearly not what people mean.

For the deeper treatment of how AGI is defined and contested, see our AGI explained guide.

Narrow AI examples in 2026

Almost every AI system in production is narrow in the strict sense, even when it does not feel narrow.

Notice the pattern: even systems built on general-purpose LLMs end up classified as narrow when the product wraps them in a specific task. The base model may be general-ish; the deployed system is narrow. This is the source of a lot of confusion in popular writing, where the underlying model and the deployed product are conflated.

Why "general" is harder to define than people think

The challenge with "general" is that humans, the reference class, are not actually general in any rigorous sense. A typical adult cannot solve a research-level mathematics problem, perform surgery, fly a helicopter, or speak Mandarin Chinese without specific training. What humans have is the ability to learn any of these given the right setup -- but learning is itself a competence that varies wildly between individuals and tasks.

This produces a definitional pickle. If "general" means "performs every task a human can", almost no human qualifies. If it means "can learn any task a human can", you have to specify the data, the time, and the setup, at which point the question becomes how the system learns rather than what it is. If it means "performs most economically valuable tasks at human level", you have something measurable but you have moved the goalposts substantially -- this is closer to the OpenAI working definition.

The deeper problem is that intelligence in the human sense is bundled with embodiment, social context, motivation, and a lifetime of physical interaction with the world. Removing those and asking whether a disembodied text model is "general" is asking a different question than the original taxonomy intended. Some researchers (LeCun, Mitchell, Marcus) have argued that the question is incoherent without grounding the model in something more than text. Others (the OpenAI and Anthropic camps) treat capability across a wide enough range of tasks as evidence of generality regardless of the cognitive architecture. The disagreement is unresolved as of 2026.

Reactive, limited-memory, theory-of-mind, self-aware

The older AI textbook taxonomy, popularised by Arend Hintze in 2016 and still cited widely, has four types. They are organised by cognitive sophistication rather than capability scope.

Type 1 -- Reactive machines. Systems with no memory and no model of past states. They respond to current input only. IBM's Deep Blue, the chess computer that beat Kasparov in 1997, is the canonical example: it computes the best move from the current board state without any memory of how the game got there.

Type 2 -- Limited memory. Systems that use a bounded window of past information to inform current decisions. Self-driving cars belong here -- they integrate sensor history over the past few seconds to track other vehicles and predict trajectories. Almost all modern deep learning systems are Type 2 in some form, with their context window or replay buffer serving as the memory.

Type 3 -- Theory of mind. Systems that model the mental states of other agents. As of 2026, no production system reliably does this, although LLMs show partial theory-of-mind behaviour on some benchmarks (the Sally-Anne false-belief task is the classic). Whether this counts as a real model of other minds or as pattern matching from training data is contested.

Type 4 -- Self-aware. Systems with a model of themselves as agents in the world. Speculative; no current system qualifies. The closest current behaviour is LLMs that report on their own reasoning steps via chain-of-thought, which is a procedural artefact, not self-awareness in any deep sense.

This taxonomy is more philosophically rigorous than the narrow/general/super one, but less useful for talking about what current systems can do commercially. Both schemes coexist in the field; you will see both in textbooks and articles.

The 2026 state -- where current models sit

If you take a frontier model in 2026 (GPT-5, Claude 4 Opus, Gemini 2.5 Pro) and place it on both taxonomies, the results are interesting and uncomfortable.

On the narrow/general/super axis: not narrow in the historical sense, since a single model handles thousands of tasks; not general in the strong sense, since it fails predictably outside its training distribution and cannot learn new skills permanently from a single example; not super by any reasonable test, since humans still outperform it on long-horizon planning, novel reasoning, and most physical tasks. The honest label is something like "general-leaning narrow" or "competent generalist", which is exactly the territory the taxonomy was supposed to make easy to describe and ends up making confusing.

On the four-types axis: clearly Type 2 (limited memory), with the context window serving as the memory. Borderline Type 3 (theory of mind) on benchmarks, but this is fragile and breaks under careful probing. Not Type 4 by any measure.

The right summary is that current systems exceed the original conception of "narrow AI" while falling well short of any rigorous "general" or "self-aware" threshold. The taxonomies were designed for systems that no longer exist (purely task-specific narrow AI) and systems that do not yet exist (true AGI), with little vocabulary for what we actually have. Several researchers have proposed updated taxonomies; DeepMind's 2023 levels paper (Emerging / Competent / Expert / Virtuoso / Superhuman, crossed with Narrow / General) is the most useful in current circulation.

For the deeper history of how the field arrived here, see our history of AI. For the active discussion about timelines and definitions, see AGI explained.

Other useful taxonomies

Beyond the two main schemes, several other ways of dividing AI systems are worth knowing because they show up in different contexts.

By learning style. Supervised learning, unsupervised learning, reinforcement learning, self-supervised learning, semi-supervised learning. This taxonomy is about how the system was trained, not what it can do. Useful for technical reading; less useful for product evaluation. Covered in our how ML and DL work guide.

By modality. Text AI, vision AI, audio AI, multi-modal AI. This taxonomy is about input and output types. It mattered more when models were single-modality; in 2026 most frontier systems are multi-modal by default and the categorisation is increasingly historical.

By symbolic vs sub-symbolic. Symbolic AI (rules, logic, search) vs sub-symbolic AI (neural networks, statistical methods). This is the oldest big-picture taxonomy and still useful in academic contexts. Most current AI is sub-symbolic; neuro-symbolic AI tries to combine both.

By DeepMind's 2023 levels framework. Crosses two axes: capability level (Emerging / Competent / Expert / Virtuoso / Superhuman) and generality (Narrow vs General). Produces ten cells, each describing a meaningful state of progress. Useful for talking about specific systems with precision; the most-rigorous current alternative to the narrow/general/super ladder.

By risk tier (regulatory). The EU AI Act's classification: prohibited, high-risk, limited-risk, minimal-risk. Specifically not about technical capability; about context of use. Increasingly important for compliance and governance. Covered in our ethics guide.

The right taxonomy depends on the question you are trying to answer. Capability scope, cognitive style, learning method, modality, technique family, regulatory risk -- each axis is useful in some contexts. The mistake is treating any one as the canonical answer.

Frequently asked questions

How many types of AI are there?

It depends on the taxonomy. The capability ladder has three (narrow, general, super). The cognitive-style taxonomy has four (reactive, limited-memory, theory-of-mind, self-aware). DeepMind's 2023 framework crosses two axes (level of competence x narrow vs general) and produces ten cells. Each is useful for a different conversation. There is no single canonical answer.

Are large language models AGI?

Not by any rigorous definition. Frontier LLMs in 2026 are general across many tasks but fail predictably on long-horizon planning, novel reasoning, and reliable self-correction. They satisfy some weak definitions of generality and fail strong ones. Whether they will scale into AGI is the central open question of the field. See AGI explained for the longer version.

What is the difference between weak AI and strong AI?

"Weak AI" and "strong AI" are older terms, originally from John Searle's 1980 Chinese Room paper, where strong AI meant a system that genuinely understands and weak AI meant a system that simulates understanding without really having it. In modern casual usage, "weak AI" is sometimes used for narrow AI and "strong AI" for AGI, which is a different question than Searle's original one. The modern usage is sloppy but common.

Is reactive AI obsolete?

Not entirely. Reactive systems are still appropriate where the task genuinely does not need memory: real-time signal processing, certain kinds of game-playing, simple control loops. But for almost any task involving language, reasoning, or interaction with humans, you want at least limited memory. Most reactive systems in production have been replaced by limited-memory systems over the past decade.

Will we know when we have built AGI?

Probably not cleanly. The most likely pattern is gradual capability accretion that crosses different thresholds at different times for different observers. Some researchers will declare AGI when an arbitrary benchmark suite is saturated. Others will hold out for behaviour that frontier models currently fail at. Expect a contested, drawn-out transition rather than a single moment.

What about superintelligence -- is that even possible?

Whether ASI is possible depends on assumptions that are not currently testable. The intuition that an AGI would quickly become an ASI rests on the ability of the system to improve itself, which in turn depends on the long-horizon planning ability current models lack. If those improve, the gap between AGI and ASI may be small in calendar time. If they hit a wall, the gap may be large or impossible to cross. Serious researchers disagree about which it will be.

The bottom line

The taxonomies of AI are vocabulary, not ground truth. They are useful for organising conversations and orienting yourself, but the systems built in 2026 do not sit cleanly inside any of the categories the taxonomies define. Treat narrow / general / super as a rough scope axis. Treat reactive / limited-memory / theory-of-mind / self-aware as a rough cognitive-style axis. Use both as background; do not let either tell you what the system you are evaluating actually does. The right way to evaluate any specific AI system is to look at its benchmark performance, its failure modes, and the workflow it is being deployed into -- not at which type of AI you think it belongs to. The labels are noise; the behaviour is signal.

Last updated: May 2026