What are the types of AI?
There are three common ways to sort AI, and each one answers a different question: how capable the system is, what it remembers while it works, or the technique its builders used to make it. Pick the question first.

| Question | Way to sort | Types | Exists today? |
|---|---|---|---|
| How capable is it? | By capability | Narrow, general, super | Only narrow, per IBM |
| What does it remember? | By function | Reactive machines, limited memory, theory of mind, self aware | The first two, per Hintze |
| How was it built? | By technique | Rules, machine learning, deep learning, generative | All four |
People often merge the three lists into one, and that is where the claims of "7 types of AI" come from. Seven is just three capability types plus four function types. It isn't a seventh kind of machine. The types of artificial intelligence are one idea seen from three sides.
The dates above come from Wikipedia on MYCIN, IBM on Deep Blue and The Conversation on Hintze. They show why the old labels fit old machines best. Two of the four function types were drawn up in 2016, before ChatGPT existed.
What are the three types of AI by capability?
By capability, the three types are narrow AI, general AI and super AI, and only the first one exists. IBM says narrow AI "is the only type of AI that exists today" and that it "can't perform outside of its defined task".
Narrow AI (also called weak AI or ANI) covers every tool you can use now: a chess engine, a spam filter, a voice assistant, a chatbot. IBM names Siri, Alexa, Watson and ChatGPT as examples. A chatbot feels broad because it writes, codes and translates, yet IBM still files it here, calling ChatGPT "limited to the single task of text-based chat". IBM also places reactive machines and limited memory AI inside narrow AI, which is how the capability and function lists connect.
General AI (AGI or strong AI) would learn any intellectual task a person can. IBM calls it "today nothing more than a theoretical concept". Super AI (ASI) would think and reason beyond human ability, and IBM calls it "strictly theoretical".
The sources split on one point. Google DeepMind's Levels of AGI paper treats generality as a ladder, with chatbots on its first rung. Our AGI vs AI page gives both lab definitions in full, so we won't repeat them here.
What are the four types of AI by function?
By function, the four types are reactive machines, limited memory, theory of mind and self aware AI. Arend Hintze, then an assistant professor at Michigan State University, laid them out on The Conversation on 13 November 2016.
| Type | Hintze's idea | Real example | Exists? |
|---|---|---|---|
| Reactive machines | "purely reactive", with no memories and no use of past experience | IBM's Deep Blue, May 1997 | Yes |
| Limited memory | Can "look into the past" but keeps it only briefly | Self driving cars that track other cars' speed and direction | Yes |
| Theory of mind | Understands that people and things can "have thoughts and emotions" | None named by Hintze | Not yet |
| Self awareness | Systems that "form representations about themselves" | None named by Hintze | Not yet |
Deep Blue is the textbook reactive machine, and Hintze names AlphaGo as a second one. Deep Blue looked at the board in front of it and picked a move. IBM's own page says it beat Garry Kasparov 3.5 to 2.5 in May 1997, evaluating 200 million positions a second.
Limited memory is the type that causes the arguments. Hintze's example is a self driving car that tracks the speed and direction of nearby cars for a short time, then drops it. Coursera repeats his four types and says self aware AI doesn't exist yet.
What type of AI is ChatGPT?
ChatGPT is narrow AI by capability, and a generative, deep learning system by technique. Its function type is where sources disagree, because Hintze wrote his scheme before chatbots like it existed.
IBM files ChatGPT under narrow AI. Future Skills Academy files ChatGPT and Gemini under limited memory, and Coursera treats ChatGPT as generative AI that sits outside the four function types. Three sources, three answers.
Our placement is the cautious one. A chat model keeps what you typed in its context window for the length of the conversation, which looks like limited memory, and it isn't reactive because it uses earlier turns to shape each new reply. Nothing in these sources supports theory of mind or self awareness. Treat any single label as a convenience, not a fact about the model.
What are the types of AI by technique?
By technique, the types are rule based systems, machine learning, deep learning and generative AI, and they nest more than they compete. IBM describes AI as the broad field, machine learning as a subset of it, and deep learning as a subfield of machine learning.
- Rule based AI. A person writes the rules and the program applies them. Wikipedia describes an expert system as "a computer system emulating the decision-making ability of a human expert", with a knowledge base of facts and rules and an inference engine that applies them. Stanford's MYCIN, an infection diagnosis system, published its reasoning model in 1975.
- Machine learning. The program learns patterns from data instead of following written rules. Spam filters and recommendation engines work this way. VidHex lists its three styles as supervised learning (labelled data), unsupervised learning (unlabelled data) and reinforcement learning (trial and error). What is machine learning covers the three main styles.
- Deep learning. Machine learning that uses neural networks with many layers. IBM says "deep" refers to the number of layers, and that deep learning generally needs more data than classic machine learning.
- Generative AI. IBM defines it as AI "that can create original content such as text, images, video, audio or software code". It names ChatGPT, DALL-E, Midjourney and Stable Diffusion. See what is generative AI for how it works.
This is the lens that matches how products get built today. A voice assistant can use rules for a timer, machine learning to hear you and a generative model to answer. One product, three techniques.
Two other fields get filed beside these. Blue Prism lists natural language processing (human language) and computer vision (images and video). Blue Prism also sorts AI by business use, naming generative AI (creates new content), predictive AI (estimates what happens next) and agentic AI (pursues goals across multiple steps). That is a fourth question, what the system is for, and it overlaps the technique list. For the last one, see agentic AI vs generative AI.
What are the types of AI models?
The types of AI models follow the technique list, and the useful split today is by what a model produces. Text models write and answer, image models draw, and older predictive models score or sort things. IBM's generative AI page names GPT-4, Copilot and BERT for text, DALL-E, Midjourney and Stable Diffusion for images, and Meta's Llama 2 as an open source model.
A model type is not a product type. ChatGPT is a product built on language models, and a self driving car is a product built on several vision and planning models at once. When someone asks about types of AI models, ask whether they mean the model underneath or the product on top. The answers differ.
Which real systems fit which type?
Every system has three labels, and the table gives all three. The example and the capability label come from the named source. The function and technique labels are our placement, and we mark where a source disagrees.
| System | Capability | Function | Technique | Source of the example |
|---|---|---|---|---|
| Deep Blue (1997) | Narrow | Reactive machine | Search plus hand tuned rules | IBM history, Hintze |
| MYCIN (1975) | Narrow | Reactive machine (our placement) | Rule based expert system | Wikipedia |
| Self driving car | Narrow | Limited memory | Machine learning and deep learning | Hintze |
| Siri or Alexa | Narrow | Limited memory (our placement) | Machine learning, language models | IBM |
| Netflix recommendations | Narrow | Reactive machine, per Coursera | Machine learning | Coursera |
| ChatGPT | Narrow, per IBM | Disputed, see above | Deep learning, generative | IBM |
| DALL-E or Midjourney | Narrow | Limited memory (our placement) | Deep learning, generative | IBM |
| General AI, super AI | Theoretical | None exist | None exist | IBM |
Whizi loses this table to the sources themselves. Hintze's article and IBM's pages are the ones to cite, and no chatbot summary should stand in for them.
How many types of artificial intelligence are there?
Count the lists, not the machines: one list has three entries, one has four and one has four, and none of them is the official one. No standards body sets the types of AI. Each list is a teaching aid that one person published and many others copied, usually without saying which question it was built to answer.
Some pages quote seven, and others quote two (narrow and general). A page that says "there are exactly four types" is picking one scheme and not saying so. If a source doesn't say which question its list answers, capability, function or technique, don't trust its count.
What do popular guides get wrong about the types of AI?
Three things stand out in the top results we read on 22 November 2026, and each one can be checked against a primary source.
- A date for neural networks. Future Skills Academy, published 3 April 2024, says "Artificial neural networks arrived in 2012". Neural networks are far older, and Hintze's 2016 article already describes neural network evaluation in AlphaGo.
- Unsourced labels. The same guide files ChatGPT and Gemini under limited memory with no source, and IBM files ChatGPT under narrow AI. One of the two is a judgment call.
- Mixed lists. VidHex (updated 11 March 2026) has a part called "Based on an AI Model" with machine learning, deep learning and natural language processing. The first two are techniques and the third is a field. It also describes self aware AI as existing "in science fiction and discussions", while its own table calls it hypothetical. No source here reports a real self aware system, and Hintze says we are probably far from building one.
Neither guide cites a primary source for its table, and neither dates a single example. Treat them as summaries. And treat an unsourced percentage or count on any page, ours included, as something to check before you repeat it.
What can you ask an AI about the types of AI?
Paste any of these into a chatbot with web search on, and ask it to link every claim. Open the links before you trust the answer.
- "Explain the difference between narrow AI, general AI and super AI. Quote IBM's own page for each, with the link, and say not stated where the page is silent."
- "Summarize Arend Hintze's 2016 article on four types of AI in 60 words, then say what has changed since chatbots appeared."
- "Place [name of an AI product] in all three schemes: capability, function and technique. Say which placement you are unsure about and why."
- "Give me three real examples of reactive machines and three of limited memory AI, each with a source I can open."
- "List where two well known guides to the types of AI disagree about ChatGPT, and which primary source settles it."
- "Explain the types of AI to a [age] year old using one example from each scheme."
Where does Whizi fit?
Whizi is one app instead of a pile of single purpose tools, and the quickest way to see the types of AI is to ask one question to several chatbots. Pro carries Claude, GPT and Gemini, and its first week is $1.99. The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card.
Side by side Compare, on Pro and Powerhouse, runs two models on one prompt. Try prompt 3 on two of them and read where their placements differ. Web search is on by default per chat, so a prompt can fetch IBM's page instead of recalling it. Whizi loses to the primary pages for any quote.
Where do these type schemes fall short?
None of the schemes has a referee, and all of them blur at the edges. IBM's capability list calls a chatbot narrow although it handles many tasks, Hintze's function list was written before large language models existed, and the technique list overlaps because every deep learning model is also a machine learning model. Expect arguments.
Our sources have limits too. IBM's types page is dated 14 August 2024, and its generative AI page is dated 22 March 2024 and updated 2 October 2026. Hintze's article is from 2016. The Coursera and Future Skills pages are teaching summaries and cite little. For the research view of generality, read DeepMind's paper instead of any of them.
- Say which question a list of AI types answers before you count it
- Give every system a capability, a function and a technique label
- Check the date of the source, since the function types come before chatbots
- Treat a count such as 7 types as a merge of lists, not a new kind
- Quote only what you read on the original page
Frequently asked questions
What are the 4 types of AI?
The four types are reactive machines, limited memory, theory of mind and self aware AI, from Arend Hintze's 2016 article. Only the first two exist today. Deep Blue is the reactive example, and a self driving car is the limited memory example.
What are the 3 types of AI?
The three types are narrow AI, general AI and super AI. IBM says narrow AI is the only type that exists today, and that general AI is a theoretical concept. Siri, Alexa and ChatGPT are IBM's narrow examples.
Is ChatGPT narrow AI or general AI?
IBM files ChatGPT under narrow AI. Other sources put it in limited memory or in generative AI, which shows the labels aren't settled. DeepMind's paper treats generality as levels, with chatbots at the first one. See AGI vs AI.
What is the difference between generative AI and other types of AI?
Generative AI makes new content, and other types classify, predict or follow rules. IBM defines it as AI that can create text, images, video, audio or code. It is a technique, so it sits across the capability and function types.