
The types of AI are usually grouped two ways. By capability: narrow AI (ANI) that does specific tasks, artificial general intelligence (AGI) that would match humans across most tasks, and artificial superintelligence (ASI) that would exceed them. By function: reactive machines, limited memory, theory of mind and self-aware AI. Only narrow, limited-memory AI clearly exists today.
The two schemes answer different questions. Capability asks how broad and how strong the intelligence is. Function asks how the system uses memory and models other minds. This page covers both, then maps today’s tools onto them. New to the topic? Start with what is artificial intelligence.
Types of AI at a glance
| Scheme | Type | What it means | Exists today? | Example |
|---|---|---|---|---|
| Capability | Narrow AI (ANI) | Built or trained for specific tasks | Yes | Spam filters, face unlock, ChatGPT (by most accounts) |
| Capability | General AI (AGI) | Human-level across most cognitive tasks | Disputed; not by most definitions | None agreed |
| Capability | Superintelligence (ASI) | Far beyond the best humans in nearly every field | No | Hypothetical |
| Function | Reactive machines | No memory; responds only to the current input | Yes | IBM Deep Blue (chess, 1997) |
| Function | Limited memory | Uses learned or recent past data to decide | Yes | Self-driving systems, LLMs, recommenders |
| Function | Theory of mind | Models other agents’ beliefs and emotions | No (research goal) | None |
| Function | Self-aware | Represents itself and its own internal states | No | None |
What are the 3 types of AI by capability?
Narrow, general and super. These terms describe a ladder of breadth and power.
Narrow AI (ANI, “weak AI”)
AI that does one job or a bounded set of jobs. A chess engine can’t drive a car; a translation model can’t read an X-ray unless it was also trained for that. “Weak” refers to scope, not quality: narrow AI is often superhuman within its task.
Large language models complicate the picture. They write, code, summarise and reason across many domains, so they are far more general than older narrow systems. Google DeepMind’s 2023 “Levels of AGI” paper places frontier chatbots at “Emerging AGI” on its own scale, while most researchers still say they fall short of AGI because they fail unpredictably and can’t reliably finish long tasks on their own.
Artificial general intelligence (AGI, “strong AI”)
AI that can learn and do most intellectual work as well as a capable human. OpenAI’s Charter defines it as “highly autonomous systems that outperform humans at most economically valuable work.” There’s no single accepted test, so whether AGI has arrived depends on whose definition you use. Full guide: what is AGI.
Artificial superintelligence (ASI)
AI that greatly exceeds the best humans in practically every domain. Philosopher Nick Bostrom popularised the idea in Superintelligence (2014). ASI is the subject of this site: see what is super intelligence and how it differs from AGI in AGI vs ASI.
What are the 4 types of AI by functionality?
Reactive machines, limited memory, theory of mind and self-awareness. Arend Hintze, then at Michigan State University, laid out this four-step scheme in The Conversation in November 2016. It has since been repeated in many textbooks and vendor explainers.
- Reactive machines. No memory of past events; they evaluate the present situation and respond. Hintze’s example is IBM’s Deep Blue, which beat world champion Garry Kasparov in 1997 by assessing the current board.
- Limited memory. They look a short way into the past. Hintze’s example is a self-driving car tracking the speed and direction of nearby cars. Most modern machine-learning systems fit here: they learn from training data and use recent context (such as a chat history) to decide.
- Theory of mind. A future type that would understand that people and other agents have their own beliefs, intentions and emotions, and adjust accordingly. Hintze argued this is needed for AI to work socially with humans.
- Self-awareness. A hypothetical type that would form representations of itself and its own internal states. Hintze wrote that researchers are probably far from building it.
How do the two schemes fit together?
They’re independent axes. A recommender is narrow and limited-memory; a future AGI would probably need at least theory-of-mind abilities. Our view: the capability ladder is the more useful one for following the news, because the live debates (Is AGI here? When might ASI arrive?) are framed that way. The functional scheme is mainly a teaching tool; few researchers use it to classify real systems.
Which type of AI is ChatGPT?
Narrow AI with limited memory, by most accounts. ChatGPT draws on its training plus your conversation context (limited memory). Claims that it shows theory of mind in text-based tests are contested, and nothing shows it is self-aware. On the capability ladder it sits at the top of narrow AI, which is exactly why the AGI debate is so heated.
Other ways to classify AI
- By technique: rule-based (symbolic) AI vs machine learning and deep learning.
- By output: predictive AI (scores, forecasts) vs generative AI (text, images, audio, video).
- By learning method: supervised, unsupervised and reinforcement learning.
For everyday uses of each type, see our AI examples; for what the letters stand for, see AI meaning.
Enjoyed this? Get the SI Briefing
One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.
Frequently asked questions
What are the 3 types of AI?
By capability: narrow AI (ANI), which does specific tasks; artificial general intelligence (AGI), which would match humans across most tasks; and artificial superintelligence (ASI), which would far exceed humans. Only narrow AI clearly exists today.
What are the 4 types of AI?
By functionality: reactive machines, limited memory, theory of mind and self-aware AI. The scheme comes from a 2016 article by Arend Hintze. Reactive and limited-memory systems exist; the other two are hypothetical.
What is narrow AI?
Narrow AI, also called weak AI or ANI, is AI built or trained for a specific task or range of tasks, such as recognising faces, translating text or recommending videos. It cannot transfer its skill to unrelated jobs the way a person can.
Is ChatGPT narrow AI or general AI?
Most researchers still class ChatGPT as narrow AI, or at most an early, partial step toward AGI. It handles many tasks but is unreliable on long autonomous work and makes basic errors, so it does not meet most AGI definitions.
Which type of AI is most common?
Limited-memory narrow AI. Almost every system in use, from driver-assistance features to chatbots and recommendation engines, learns from past data and uses recent inputs to decide what to do.
6 Sources
- Arend Hintze, The Conversation, Understanding the four types of AI (14 Nov 2016)
- Morris et al., Google DeepMind, “Levels of AGI” (2023)
- OpenAI Charter (definition of AGI)
- Superintelligence: Paths, Dangers, Strategies (Wikipedia)
- Deep Blue versus Garry Kasparov (Wikipedia)
- Weak artificial intelligence (Wikipedia)