Types of Artificial Intelligence (AI)

Nobody taught you that you are using five different types of AI before you have your lunch.

Did you know that your email client has already filtered three different pieces of spam mail while you were sleeping? Or that your navigation app rerouted you around a car crash in real time? Or that the chatbot that answered your support question never had the chance to look at it until after the human had read it? Or that if you used any AI assistant to write anything this morning, you were talking to something that did not really exist in any meaningful way until four years ago.

The thing is, all of these are different types of AI. Not different products, different types of AI. And the way we categorize them. Well, that has largely been settled in 2019. The world has moved on.

This is your guide to the full current map – the classic taxonomy, what’s been added since, and a plain-English explanation of what actually sets one type apart from another. By the end of this article, you’ll know exactly what type is powering the tools you use every day.

The Two Ways to Categorize AI

There are two lenses through which researchers view an AI system, and they are:

The first lens is the lens of capability. How broad is the scope of the intelligence of this particular system? Is it only good for one thing, or is it good for multiple things, or is it good for anything?  

The second lens is the lens of functionality. How does the system actually work? Does it remember? Does it learn? Does it try to simulate what goes on inside another mind?  

Both are important. A self-driving car is a narrow AI(capability), but limited memory AI(functionality) and understanding both labels will help you more than understanding one. Always keep this layering in mind.

Types of AI by Capability

#1. Narrow AI — The Only Kind That Actually Exists Today

Every AI in production today—every single one of them—is narrow AI.

That includes ChatGPT, Claude, Gemini, Midjourney. The spam filter in your inbox. The AI that determines what TikTok video you’ll see next. They’re all, technically, narrow AI systems. Which sounds rather insulting until you realize that “narrow” refers to intelligence scope rather than intelligence quality.

Narrow AI—also called Weak AI—was created to do a single task or a limited range of tasks. And it does those tasks incredibly well. But outside of that range? Forget it. It completely disintegrates as an intelligent entity. Deep Blue could defeat a chess grandmaster. But it wouldn’t have been able to describe what a chess piece looked like, what it smelled like in a chess club, or even if it was worth accepting a draw.

The “weakness” is that boundary. Not the capability within that boundary.

What surprises most people – and what is obvious until it isn’t – is that generative AI tools with virtually unlimited conversation potential are still narrow AI. They are only capable of language-based interaction. Ask one to physically manipulate an object, perceive temperature, or learn a new domain from zero examples like a child does, and that boundary is immediately apparent.

#2. General AI (AGI) — What It Would Actually Take

AGI is not “smarter narrow AI.” That is completely missing the point.

The difference is not in degree but in kind. A general AI would be able to move knowledge from completely different domains without retraining. A human who learns to play chess and then learns to cook is not retrained between those two skills. They use abstract thinking, pattern recognition, and judgment between those two skills. No AI is able to do that today.

The researchers most involved in this topic are candid about how distant this is. Geoffrey Hinton, who received the Nobel Prize in Physics in 2024 along with John Hopfield for foundational contributions to neural networks, left Google in 2023 to freely discuss the existential risks of AI: “the risk of misuse by bad actors, the risk of lots of technological unemployment, and the risk that the goals programmed into an intelligence system become incompatible with human survival.” Yoshua Bengio of Mila has argued that “we might be taking a very wrong approach to building AGI with our current deep learning techniques,” and that “a major rethinking of architectures is likely necessary.” DeepMind, Anthropic, and OpenAI are all exploring different paths. None of them have arrived.

The Honest Truth – AGI is currently a purely theoretical construct. Much serious and important work is being done. The timing is genuinely unknown — from “in a decade or less” to “not this century” according to credible experts in 2026. What everyone agrees on is that we aren’t there yet.

#3. Superintelligent AI (ASI) — Why Serious Researchers Disagree

But ASI won’t only be more intelligent than human intelligence. It will be more intelligent across all dimensions—creative, rational, judicious, emotionally intelligent, and capable of figuring out tough problems. Not only will ASI be more intelligent than human intelligence, but it will also be more intelligent across all dimensions—creative, rational, judicious, emotionally intelligent, and capable of solving complex problems. Not only will ASI be more intelligent than human intelligence, but it will also be more intelligent across all dimensions—creative, rational, judicious, emotionally intelligent, and capable of solving complex problems.

The idea of ASI gained popularity for a broad audience through Nick Bostrom’s 2014 book *Superintelligence*. Bostrom’s work introduced the alignment problem that is currently at the forefront of all safety discussions related to artificial intelligence—what do you do if you have a superintelligent intelligence that is not aligned to your goals?

The alignment problem is not yet resolved. However, it is not the alignment problem for ASI, which does not yet exist. Rather, it is for narrow and agential artificial intelligence systems currently being deployed and causing harm

The debate among researchers is not about whether ASI is possible, but about whether it is inevitable, when it will happen, and whether we are developing safety systems fast enough. I think that the limitations section of any discussion about AI is more useful than the capabilities section — half the skill is knowing where it breaks.

Types of AI by How They Function

#1. Reactive AI — Pure Input-Output, No Memory

The first level of functionality is reactive machines. They are the simplest kind of machines and are capable of performing one task: receive input, process the input according to a predetermined model, and produce output.

The quintessential example of reactive machines is IBM’s Deep Blue, which beat world chess champion Garry Kasparov in 1997 by analyzing the current state of the chessboard and calculating the probabilities of the outcome of the different moves.

Deep Blue was incredibly good at what it did. However, it had no memory of previous matches. Every time the machine played, it began with the same blank slate.

The strange thing is that this makes reactive machines the most limited and the most predictable of all AI types. A machine with no memory has no way of being manipulated through its history.

#2. Limited Memory AI — How Most Modern AI Actually Works

This is where almost everything you interact with actually lives.

Limited memory AI can look at a window of past information to help generate its current output. Self-driving cars look at the movement of other cars over the past few seconds to predict the location of those cars the next few seconds. ChatGPT and Claude have a context window, a certain amount of past information that it can “see” as it is generating a message.

The word “limited” is doing some real work there.

It should be noted that the above models do not create a permanent, cumulative memory like humans do. The context of a ChatGPT session will be forgotten as soon as the session is over. It will only “remember” you the next time you give it that information, or if you have a memory feature that stores a summary of past interactions.

Most machine learning models, most large language models, most modern AI applications — all limited memory. It’s the dominant functional class of AI actually running in production today.

• How large language models (LLMs) work

#3. Theory of Mind AI — Where the Field Is Heading

The term “theory of mind” is used in developmental psychology. In the case of humans, “theory of mind” is the capacity to understand that other people have mental states like beliefs, intentions, feelings, and desires. In addition, these mental states may differ from your own. In children, this capacity starts developing when they are about four years old.

Currently, no artificial intelligence system has this capacity. However, some artificial intelligence systems are able to recognize emotional cues in text and voice. This is not “theory of mind.” This capability is merely matching patterns. Emotion AI research, conducted by institutions like MIT Media Lab and companies like Affectiva, is trying to achieve “theory of mind.” While this field of research has seen some progress, this progress is limited. 

The “theory of mind” capability that artificial intelligence will achieve will be important. This capability will allow artificial intelligence to change its actions based on the reason for your query, not merely the query.

#4. Self-Aware AI — Science Fiction or Near Future?

The short answer is: nobody genuinely knows, and anyone who tells you confidently that it’s either imminent or impossible is working from belief, not evidence.

A self-aware AI would be conscious: that is, it would have an awareness of its own existence and its place in the world. Of course, one of the biggest problems here is that nobody is really sure what consciousness is. We don’t have a definition for it. We don’t have a test that can be used to confirm or deny its existence in any given system.

What we do have is a lot of argument and hand-wringing, some very interesting science fiction, and a research community that is at least honest about how little it knows.

The standard taxonomy was defined in a period when AI systems were simpler and more siloed. In 2026, three additional types are too important to treat as footnotes.

#5 Generative AI — The Type That Changed Everything

Generative AI is more than a classifier or a predictor; it is a creator of new content, not a classifier or a decision-maker.

The technical definition of generative AI is a model that learns the underlying distribution of the training data well enough to generate new instances of that data.

The technical distinction between generative and discriminative models is that the generative model learns the underlying distribution of the training data well enough to generate new instances of the data, while discriminative models learn the boundary between classes, e.g., “is this a cat or a dog?”

The transformer architecture, as introduced in the 2017 paper “Attention Is All You Need,” is the technical foundation for nearly all generative AI models.

The major generative AI models as of early 2026 include OpenAI’s GPT series, Google’s Gemini, Anthropic’s Claude, and image generation models like Stable Diffusion and Midjourney.

Generative AI is a technical narrow form of artificial intelligence, but its impact is broader than almost any other narrow form of artificial intelligence, which is why it is a separate category in any honest taxonomy of artificial intelligence as of 2026.

#6. Agentic AI — When AI Takes Actions, Not Just Answers

This is the type that is most relevant to us today.

While conversational AI is designed to answer questions, agentic AI is designed to get things done. It has to design its actions to achieve its goal, use tools to accomplish its tasks (such as search, executing code, file management, and API use), and achieve its goal through multiple actions without human oversight for each action.

So there is a big difference here. Ask a conversational AI “research our competitors and write a summary report,” and it can write up some text for you. Ask an agentic AI to do the same thing, and it can search the web for relevant competitors, read relevant web pages, identify key information to know about competitors, and write up a report for you – all without you needing to do anything but ask.

This capability fundamentally alters the risk profile of AI deployment. A system that acts in the world—sends emails, runs code, buys things online, organizes files—poses different risks than a system that only produces text. The alignment problem that sounds abstract in ASI becomes very tangible in agentic AI.

Agentic AI is where AI is advancing fastest. And it’s almost completely absent in competitor articles even as recently as last year.

#7. Multimodal AI — When AI Processes Everything at Once

The first artificial intelligence systems were modal specialists. The vision system processed images. The language system processed text. They were different systems, different training, different deployment.

Multimodal artificial intelligence is a new type of artificial intelligence that processes multiple types of information at once. That’s to say, it processes text, images, audio, and video within one model. You can show it a picture and ask it a question about that picture. You can talk to it and it will respond in writing. It’s got all this information in one unified representational space.

GPT-4o and Gemini Ultra were early examples of this new type of artificial intelligence. The applications for this new type of artificial intelligence extend far beyond mere curiosity.

It is still a narrow artificial intelligence. Narrow artificial intelligence is artificial intelligence that is only doing one thing at a time. Now, it’s doing multiple things at once.

Side-by-Side Comparison: All 7 Types of AI

AI Type	Exists Today	Memory	Learns Over Time	Example	Primary Use Case
Narrow AI	✅ Yes	Varies	With retraining	Spam filter, ChatGPT	Single-domain tasks
Generative AI	✅ Yes	Context window	With fine-tuning	GPT, Claude, Midjourney	Content creation
Agentic AI	✅ Yes (early)	Session + tools	With retraining	AI coding agents	Multi-step autonomous tasks
Reactive AI	✅ Yes	❌ None	❌ No	IBM Deep Blue	Fixed-response systems
Limited Memory AI	✅ Yes	Short-term	Incremental	Self-driving cars	Dynamic decision-making
General AI (AGI)	❌ Theoretical	Persistent	Cross-domain	—	Any intellectual task
Superintelligent AI (ASI)	❌ Theoretical	Unknown	Unknown	—	Surpassing human cognition

Which Type of AI Do You Actually Need?

The majority of the articles tell you what kinds of AI exist. None of the articles help you decide.

Here is a simple framework to help you decide: Ask yourself four questions:

1. Is the problem you’re trying to solve well-defined, with a clear output? Yes, then you likely need narrow AI, a reactive or memory-constrained architecture.

2. Do you need to create new content? Yes, then you likely need generative AI. What kind of modality or model family you’ll use to create the output you need is the next question.

3. Does this task involve doing multiple steps on your own? If so, agentic AI. Research all workflows, all multi-step processes, anything that currently requires a human to chain together multiple tools.

4. Does the task require understanding why a person is asking, not just what they said? If so, you’re at the boundary of what current AI is good at. Theory of mind AI is what you want here; limited memory systems can give you this, but only to a certain extent.

And if you find yourself saying, “I need this to do all of these things, across any domain, without being retrained,” you’re looking for AGI. Except, alas, AGI doesn’t currently exist.

The Technologies Behind Every Type of AI

Every type of AI mentioned above is supported by one or more of these technologies. You don’t need to go in-depth into all of them; an understanding of these basic stacks puts all these types of AI into context.

1. Machine learning is the foundation. Unlike other computer programs, ML is not programmed by explicitly defining rules. Instead, it learns by analyzing data. The three major types of machine learning are supervised learning, unsupervised learning, and reinforcement learning. Weights are adjusted billions of times until machine learning matches the target.

2. Deep learning is an extension of machine learning that uses neural networks. The ‘deep’ in deep learning doesn’t mean profound; it refers to layers of artificial networks. This is the architecture that enabled image recognition, speech recognition, and language generation. The moment most researchers point to as marking the beginning of deep learning is the AlexNet paper by Krizhevsky, Sutskever, and Hinton in 2012.

3. The transformer architecture, which was introduced in 2017, is the particular form of neural network that underlies all of the major language models currently in use.

4. The attention mechanism, which is the particular innovation that makes the transformer work, allows the model to weigh the relative importance of different parts of the input to each other.The tokens are given numerical weights and processed in parallel, which is why the transformer was so successful.

5. Natural language processing (NLP) is the branch of computer science that allows us to work with human language. It is the NLP that has evolved to be transformer-based.

6. Reinforcement Learning from Human Feedback (RLHF) is the method by which large language models are aligned to human preferences. It is the method by which the large language models were aligned to human preferences, from “technically coherent but bizarre” to “actually useful.”

Frequently Asked Questions

Conclusion

The taxonomy of AI that you were taught in school – if you were taught one – has three classes and four classes of functionality. This was true in its time.

The AI that you’re using in 2026 does not fit into those boxes anymore. Agentic AI has reactive logic, memory, and generation in one pipe. Multimodal AI has the properties of vision AI and language AI. The boxes are merging into one another as compound capability is added.

The best way to think about AI isn’t to memorize a taxonomy. It’s to develop the instincts to understand that question: What is this system actually doing with my input? What are its actual limits?

The person who understands that question – regardless of the taxonomy – is not easily fooled by either hype and fear. And in 2026 – and in every tool you use – that’s more important than the taxonomy.