Guide
Types of AI
A plain-English map of the types of artificial intelligence: what AI is, narrow vs general vs super AI, the four functional types from reactive machines to self-aware AI, the main techniques (rule-based, machine learning, deep learning), generative vs predictive AI, the risk tiers in EU law, and where ChatGPT and other everyday tools fit.
Quick answer: There are three main types of AI by capability: narrow AI, which is built for specific tasks and covers almost all AI in everyday use; artificial general intelligence (AGI), a system that matches humans across most cognitive work, whose arrival is openly disputed; and artificial superintelligence (ASI), a system that far exceeds the best humans at almost everything, which nobody claims exists. A second, widely taught scheme sorts AI by how it functions: reactive machines, limited memory, theory of mind and self-aware AI, from a 2016 article by Michigan State University’s Arend Hintze. ChatGPT, Claude and Gemini are generative AI built on large language models, and whether the newest of them count as general AI is the live argument. The one caveat: these schemes are teaching frameworks, not official standards, so two articles can count “the types of AI” differently and both be right.
This page is the map. It explains each way of dividing up AI and links down to the deep dives: what is AGI, what is an LLM, what is agentic AI and what is RAG. For one-line definitions of everything else, see the AI glossary. For which models and apps are actually best, see best AI models and best AI chatbots.
What is artificial intelligence?
Artificial intelligence (AI) is the field of building computer systems that do tasks normally associated with human intelligence, such as understanding language, recognising images, making predictions, making decisions and creating content. An AI system learns or infers how to do those tasks from data and rules, rather than following only step-by-step instructions written by a programmer.
A widely used formal definition comes from the OECD, whose member countries revised it in 2023: “An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.” The OECD adds that “Different AI systems vary in their levels of autonomy and adaptiveness after deployment.” The EU’s AI Act, Regulation (EU) 2024/1689, uses a closely matching definition in Article 3(1).
The key word in both definitions is infers. A calculator follows fixed rules and is not AI. A spam filter that has learned from millions of emails which messages look like spam is AI, because it works out for itself how to turn an input into an output.
The term itself dates from a proposal for the Dartmouth summer research project dated 31 August 1955 and signed by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon.
The four ways to classify AI
People mean different things by “types of AI” because there are four separate ways to divide the field. Each answers a different question.
| Classification | The question it answers | The types |
|---|---|---|
| By capability | How broad and how strong is it? | Narrow AI, general AI (AGI), superintelligence (ASI) |
| By functionality | What can it remember and understand? | Reactive machines, limited memory, theory of mind, self-aware |
| By technique | How is it built? | Rule-based (symbolic) AI, machine learning, deep learning |
| By output | What does it do? | Predictive AI, generative AI, agentic AI, plus perception and robotics |
The rest of this page takes each one in turn, then shows where familiar products sit.
Types of AI by capability: narrow, general and super
The capability scheme is the best known. It sorts systems by how wide a range of tasks they handle and how well.
| Type | Also called | What it means | Does it exist? |
|---|---|---|---|
| Narrow AI | Weak AI, artificial narrow intelligence (ANI) | Competent at a specific task or set of tasks | Yes: almost every AI product in use |
| General AI | Strong AI, artificial general intelligence (AGI) | Matches or exceeds humans across most cognitive tasks | Disputed: some industry leaders say yes, many researchers say not yet |
| Superintelligence | Artificial superintelligence (ASI) | Far exceeds the best humans in virtually every domain | No: several companies name it as a goal |
Narrow AI
Narrow AI is AI built for, and competent at, a specific task or domain. A chess engine, a spam filter, a voice assistant, a fraud-detection system and a protein-structure predictor are all narrow AI. Some are far better than any human at their one job: Google DeepMind’s Levels of AGI paper lists AlphaFold, AlphaZero and the Stockfish chess engine as “Superhuman” narrow AI, and IBM’s Deep Blue and DeepMind’s AlphaGo one level below.
Narrow does not mean simple or weak. It means the system cannot carry its skill into an unrelated task. AlphaGo, which beat Go champion Lee Sedol 4–1 in March 2016, cannot write an email.
Artificial general intelligence (AGI)
AGI is a hypothetical AI system that can match or exceed human ability across most or all cognitive tasks, including tasks it was not built for. There is no single agreed definition. OpenAI’s Charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work”; Google DeepMind grades it on a five-level scale.
Whether AGI has arrived is a live argument. OpenAI President Greg Brockman greeted the launch of GPT-6 Astra with “Welcome to the AGI era” (Axios), and Mike Knoop, co-founder of the ARC Prize Foundation, whose benchmark OpenAI cited, said the foundation lacked “evidence to call this AGI yet” (The Next Web). Our AGI guide sets out the competing definitions, benchmarks and claims, each attributed to whoever made it.
Artificial superintelligence (ASI)
Artificial superintelligence is AI that substantially exceeds the best humans across essentially all cognitive domains. The philosopher Nick Bostrom’s book Superintelligence defines it as “any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest”. DeepMind’s Levels of AGI paper places ASI at its top level, “Superhuman” general AI.
Nobody claims ASI exists, but it has become a stated corporate goal. Former OpenAI chief scientist Ilya Sutskever co-founded Safe Superintelligence Inc., and Meta chief executive Mark Zuckerberg announced Meta Superintelligence Labs. Timelines for ASI are forecasts, not facts, and they vary by decades depending on who is asked. For more, see what is ASI.
DeepMind’s Levels of AGI: a finer-grained version
The three-way split hides a lot. Google DeepMind’s Levels of AGI paper grades AI on two axes at once, how well it performs and how general it is, across five levels: Emerging, Competent, Expert, Exceptional and Superhuman. Its value is the framework: it shows that a system can be superhuman and narrow at the same time, and that “general” is a matter of degree rather than a switch that flips. Google DeepMind has not published a re-grading of current models, so any placement of one on the scale is a third party’s interpretation.
Types of AI by functionality: the four types
The functional scheme sorts AI by what it can remember and understand. It was popularised by Arend Hintze, then an assistant professor at Michigan State University, in a 2016 article for The Conversation, and it is the version many courses and explainers teach as “the four types of AI”.
| Type | What it can do | Examples | Exists? |
|---|---|---|---|
| 1. Reactive machines | Responds to the current input only; keeps no memory of the past | IBM’s Deep Blue, which beat Garry Kasparov in 1997 | Yes |
| 2. Limited memory | Uses recent observations to inform decisions, without building lasting experience | Self-driving cars tracking nearby vehicles; chatbots that use the current conversation | Yes |
| 3. Theory of mind | Understands that other people have their own thoughts, feelings and intentions, and adjusts to them | None agreed | No |
| 4. Self-aware | Has a model of itself and its own internal states, as conscious beings do | None | No |
Reactive machines
A reactive machine responds only to what is in front of it. Hintze calls them “the most basic types of AI systems”. Deep Blue evaluated the chess position on the board and chose a move; it did not remember earlier games or learn from them during play.
Limited memory
Limited-memory AI uses recent data to make decisions. Hintze’s example is the self-driving car, which tracks the speed and direction of nearby vehicles over a few seconds. In his framing, these “simple pieces of information about the past are only transient”: the car does not turn them into lasting experience.
Most modern AI falls into this type. A chatbot that holds the current conversation in its context window, then forgets it when the session ends, is limited-memory AI. Memory features in apps such as ChatGPT work by saving notes and feeding them back into the conversation, not by the underlying model learning from each chat. The LLM guide explains how that works.
Theory of mind
Theory-of-mind AI would understand that people have beliefs, emotions and intentions of their own, and would change its behaviour in response. The term comes from psychology, where it describes a skill children develop in early childhood. Chatbots can discuss other people’s feelings fluently, and researchers disagree about whether that reflects understanding or pattern-matching on text that describes human behaviour. No system is generally accepted as having theory of mind.
Self-aware AI
Self-aware AI would have a sense of itself: knowledge of its own internal states, and arguably consciousness. Hintze wrote in 2016 that researchers are “probably far from creating machines that are self-aware”. Whether any current or future AI system could be conscious is an open question in philosophy and neuroscience, not a solved engineering problem.
Is the four-types scheme still useful?
The four-types scheme is a good teaching tool, but it was written before modern chatbots, and it does not map neatly onto them. A large language model is limited memory by the scheme’s definition, yet it can discuss other people’s feelings fluently. The capability scheme (narrow, general, super) and the technique scheme below do more work in current discussion.
Types of AI by technique: how AI is built
The technique scheme sorts AI by how it gets its abilities. The three main families are nested: deep learning is a kind of machine learning, and machine learning is one approach to AI.
| Technique | How it works | Strengths | Examples |
|---|---|---|---|
| Rule-based (symbolic) AI | People write explicit rules and logic | Predictable, explainable, needs no training data | Expert systems, early chatbots |
| Machine learning | The system learns patterns from data instead of following hand-written rules | Handles messy, real-world data at scale | Spam filters, credit scoring, recommendation engines |
| Deep learning | Machine learning using neural networks with many layers | Excels at images, speech and language | Image recognition, speech-to-text, large language models |
Rule-based AI
Rule-based AI, sometimes called symbolic AI or good old-fashioned AI (GOFAI), was the dominant approach from the 1950s to the 1980s. People encode knowledge as rules (“if the temperature is above X and pressure is below Y, flag a fault”). It is still widely used where decisions must be predictable and auditable, but it breaks down on tasks that are hard to describe as rules, such as recognising a face. The EU AI Act draws the line with ordinary software in Recital 12: “logic- and knowledge-based approaches” count as AI, but systems running “rules defined solely by natural persons” do not, which is why an expert system can be AI and a calculator is not.
Machine learning
Machine learning is the approach in which a system learns to perform a task from examples rather than being explicitly programmed. There are three main ways to train it:
| Training method | How it learns | Typical use |
|---|---|---|
| Supervised learning | From labelled examples: emails marked “spam” or “not spam” | Classification and prediction |
| Unsupervised learning | Finds structure in unlabelled data on its own | Customer segmentation, anomaly detection |
| Reinforcement learning | Tries actions and is rewarded or penalised for the results | Game-playing systems such as AlphaGo; training reasoning models |
Large language models add a fourth, self-supervised learning: the model generates its own training signal by predicting the next token of ordinary text, so it needs no human labelling to learn from trillions of words. It is then refined with supervised examples and reinforcement learning, as the LLM guide describes.
Deep learning and neural networks
Deep learning is machine learning that uses artificial neural networks with many layers, loosely inspired by the brain. Each layer transforms its input a little, and stacking many layers lets the network learn complex patterns such as faces, speech or meaning in text. Almost every headline AI system of the last decade is deep learning, including image generators and large language models. Most large language models use the transformer, a neural-network design that Google researchers introduced in June 2017.
Types of AI by output: predictive, generative and agentic
The output scheme sorts AI by what it produces. This is the split most relevant to choosing a product.
| Type | What it produces | Examples | Related pages |
|---|---|---|---|
| Predictive AI | Scores, classifications, forecasts and recommendations | Fraud detection, demand forecasting, recommendation feeds, spam filters | Best AI for data analysis |
| Generative AI | New content: text, images, audio, video and code | ChatGPT, Claude, image and video generators | Best AI chatbots, best AI image generator, best AI video generator |
| Agentic AI | Actions: it plans, uses tools and works through multi-step tasks | Coding agents, browser agents, workflow agents | What is agentic AI, best AI agents |
| Perception AI | Understanding of images, video and sound | Face unlock, medical-image analysis, speech recognition | Best AI for transcription |
| Robotics and embodied AI | Physical movement in the real world | Self-driving cars, warehouse robots, humanoid robots | Not covered on this site |
Predictive AI vs generative AI
Predictive AI answers “what is likely?”; generative AI answers “what could this look like?”. A bank’s fraud model looks at a transaction and outputs a probability that it is fraudulent. A chatbot looks at a prompt and writes a new answer. Both are machine learning, often deep learning, and the line between them is blurring: a large language model is, mechanically, a predictor of the next token, used to generate.
Generative AI is the type behind the current AI boom. Text generators are built on large language models; image and video generators are mostly built on diffusion models, which learn to turn random noise into an image step by step. Many products are now multimodal, accepting and producing several kinds of content in one model.
Agentic AI
Agentic AI is a generative model, usually a large language model, placed in a loop with tools such as a browser, a code runner or a file system, so it can take actions, check the results and keep going until a task is done. It is less a new kind of model than a new way of using one. Our agentic AI guide covers how agents work and where they fail; best AI agents compares the products.
Types of AI in law: the EU AI Act’s risk tiers
Regulators use a fifth classification: by risk. The EU AI Act, which entered into force on 1 August 2024, sorts AI systems into four risk levels, with heavier obligations the higher the risk.
| Risk level | What it means | Examples |
|---|---|---|
| Unacceptable | Banned outright | Social scoring, manipulative techniques that exploit vulnerabilities, scraping faces to build recognition databases |
| High | Strict obligations before a system can go on the market | AI used in critical infrastructure, education, employment and law enforcement |
| Limited | Transparency duties, such as telling people they are talking to an AI | Customer-service chatbots |
| Minimal | Largely unregulated | Spam filters, AI in video games |
The bans on unacceptable-risk systems applied from 2 February 2025, and breaking them carries fines of up to €35 million or 7% of worldwide annual turnover. The Act also has separate rules for general-purpose AI models, the large models behind chatbots, because one model can end up in products at every risk level.
Where everyday AI tools fit
Most real products sit in several categories at once. This table places familiar systems on each scheme. The placements are our reading of each scheme.
| System | Capability | Functionality | Technique | Output |
|---|---|---|---|---|
| ChatGPT, Claude, Gemini | Contested: commonly called narrow, called AGI by some industry leaders | Limited memory | Deep learning (large language model) | Generative, increasingly agentic |
| Image generators such as Midjourney | Narrow | Reactive | Deep learning (diffusion model) | Generative |
| Self-driving cars | Narrow | Limited memory | Deep learning plus rules | Perception, robotics |
| Netflix-style recommendations | Narrow | Limited memory | Machine learning | Predictive |
| Email spam filter | Narrow | Reactive | Machine learning | Predictive |
| Deep Blue (chess, 1997) | Narrow | Reactive | Rule-based search | Decisions |
| AlphaFold | Narrow (“Superhuman” on DeepMind’s scale) | Reactive | Deep learning | Predictive (protein structures) |
The pattern: nearly everything here is narrow, nothing sits above limited memory, most of it is machine learning, and the products that dominate the news are generative. The disagreement is only about whether the most capable chatbots have crossed into general AI, and that depends on whose definition you use.
The Turing test: the oldest test of machine intelligence
Alan Turing proposed testing machine intelligence by conversation in his 1950 paper “Computing Machinery and Intelligence”, published in the journal Mind. In what he called the imitation game, a judge holds text conversations with a human and a machine and tries to tell which is which.
Large language models have passed controlled versions of it. In a study of five-minute conversations, University of California San Diego researchers Cameron Jones and Benjamin Bergen reported that, “When prompted to adopt a humanlike persona, GPT-4.5 was judged to be the human 73% of the time”, more often than the real human participants. The result shows that a chatbot given the right prompt can imitate a person for five minutes. It does not establish general intelligence, and current AGI claims are argued over other benchmarks.
Frequently asked questions
What are the types of AI?
The most common answer is three types by capability: narrow AI, which handles specific tasks; artificial general intelligence (AGI), which would match humans across most cognitive work; and artificial superintelligence (ASI), which would far exceed humans at almost everything. A second common answer is four types by functionality: reactive machines, limited memory, theory of mind and self-aware AI. Only narrow AI, reactive machines and limited-memory AI are generally accepted as existing.
What are the 4 types of AI?
The four types of AI are reactive machines, limited memory, theory of mind and self-aware AI, a scheme popularised by Michigan State University’s Arend Hintze in 2016. Reactive machines, such as IBM’s Deep Blue, respond only to the current input. Limited-memory systems, such as self-driving cars and chatbots, use recent information. Theory-of-mind and self-aware AI are hypothetical.
What are the 3 types of AI?
The three types of AI are narrow AI (also called weak AI), artificial general intelligence (AGI, or strong AI) and artificial superintelligence (ASI). Almost every AI product in use is narrow AI. Whether the most advanced chatbots count as AGI is disputed, and no one claims superintelligence exists.
What is artificial intelligence in simple terms?
Artificial intelligence is technology that lets computers do tasks that normally need human intelligence, such as understanding language, recognising images, making predictions and creating content. Instead of following only fixed instructions, an AI system works out from data how to turn an input into a useful output.
What type of AI is ChatGPT?
ChatGPT is generative AI built on large language models, which are a form of deep learning. By the four-types scheme it is limited-memory AI. By capability its label is contested: it is commonly called narrow AI, while OpenAI President Greg Brockman greeted the GPT-6 Astra model with “Welcome to the AGI era”, a claim others dispute.
What is the difference between narrow AI and general AI?
Narrow AI is built for a specific task or domain and cannot carry its skill into unrelated tasks, even when it is superhuman at its own job. General AI, or AGI, would perform well across most cognitive tasks, including ones it was not built for, the way a capable human can. Almost all AI products in use are narrow; whether the most capable chatbots count as general AI is debated.
What is the difference between AGI and ASI?
AGI is AI that matches or exceeds human ability across most cognitive tasks. ASI, or artificial superintelligence, is AI that far exceeds the best humans across virtually every domain. AGI is a threshold of human-level generality; ASI is well beyond it. No one claims ASI exists, while the arrival of AGI is contested.
Is generative AI a type of AI?
Generative AI is a type of AI classified by its output: it creates new content such as text, images, audio, video and code, rather than only scoring or predicting. ChatGPT, Claude, Gemini and image generators such as Midjourney are generative AI. It sits alongside predictive AI, which outputs forecasts and classifications, and agentic AI, which takes actions.
What is the difference between AI, machine learning and deep learning?
AI is the whole field of building systems that perform tasks associated with human intelligence. Machine learning is one approach to AI, in which systems learn from data instead of following hand-written rules. Deep learning is a kind of machine learning that uses neural networks with many layers, and it powers image recognition, speech recognition and large language models.
Does self-aware AI exist?
No AI system is generally accepted as self-aware. Self-aware AI would have a model of itself and its own internal states, and whether any machine could be conscious is an open question in philosophy and neuroscience. Chatbots can talk about themselves fluently, but that is not accepted as evidence of self-awareness.
What type of AI is a self-driving car?
A self-driving car is narrow, limited-memory AI. It uses deep learning to recognise roads, vehicles and pedestrians, and it tracks nearby objects over a few seconds to decide how to drive, but it cannot apply that skill to unrelated tasks. It combines perception AI with robotics.
Has AI passed the Turing test?
Large language models have passed controlled versions of the Turing test. In a study of five-minute conversations, UC San Diego researchers reported that GPT-4.5, prompted to adopt a humanlike persona, was judged to be the human 73% of the time, more often than the real people it was compared with. That shows convincing conversation, not general intelligence.