Types of AI: Narrow AI, General AI & Superintelligence

There isn’t one universally accepted list of AI types. The most common framework classifies AI by capability into Narrow AI, Artificial General Intelligence (AGI), and Artificial Superintelligence (ASI). Another commonly used framework classifies AI by functionality as Reactive Machines, Limited Memory, Theory of Mind, and Self-Aware AI.

It makes sense why people get mixed up. Most articles treat every software tool and tech buzzword as its own separate category, which turns the topic into a mess.

In reality, computer scientists sort the types of AI using two entirely different measuring sticks. One looks at capability, basically capability, basically how smart the machine is compared to a person, which gives us Narrow, General, and Super AI. The other looks at how the software actually runs under the hood, tracking how it uses memory from basic reactive machines up to systems with self-awareness.

Once you separate those two ideas, the whole landscape gets a lot easier to read.

Here is how both systems actually work, along with where everyday tools like Siri and OpenAI models land on the map. Most of them sit way lower on the ladder than the marketing wants you to believe.

Key Takeaways

•  AI is sorted two different ways: by capability (Narrow, General, Super) and by functionality (Reactive, Limited Memory, Theory of Mind, Self-Aware). That’s why the “how many types” number keeps changing by source.
• Every AI product you can actually use right now, Siri, ChatGPT, Alexa, Netflix, all of it, is Narrow AI. General AI and Super AI are still research goals, not products.
• Narrow doesn’t mean “one task.” A tool like ChatGPT can write, translate, and code, and it’s still Narrow AI because it lacks the flexible reasoning a person has.
• Terms like generative AI, agentic AI, and computer vision aren’t a separate type system. They describe what a tool does, and they sit on top of the two frameworks above.
• One system can carry several labels at once. ChatGPT is Narrow AI, Limited Memory, deep learning-based, generative, and conversational, all at the same time. 

How Many Types of AI Are There?

To understand the different types, it helps to first look at the broader field of artificial intelligence solutions and how AI systems are classified. 

Sort AI by capability, meaning how much a system can do compared with a person, and you get three types: Narrow AI, General AI, and Super AI. A commonly used conceptual framework groups AI by functionality into Reactive Machines, Limited Memory, Theory of Mind, and Self-Aware AI. Unlike the capability-based categories, however, this framework isn’t a universal industry standard, and the boundaries can vary by source.  IBM’s guide to AI types uses these same two lenses, and so do most solid explainers.

Then there’s a third pile of labels you’ll see everywhere now, like generative AI and computer vision. Those describe what a tool is used for. They aren’t a rival set of types.

Types of AI by Capability

Narrow AI (ANI)

Narrow AI, also called Weak AI, is built for a particular kind of work. It can be excellent at that work and still fall apart outside it. A chess engine can’t plan your holiday. A spam filter can block junk mail, but it can’t drive a car.

People say narrow AI does one task. That isn’t really true anymore. A chatbot like ChatGPT can write an email, summarize a report, translate a paragraph, and help debug code, and it’s still Narrow AI. The word “narrow” doesn’t mean “one task.” It means the system lacks the flexible, human-style understanding that lets a person pick up almost anything. It’s strong across a big set of jobs, not across everything.

Narrow AI is the only kind that exists in real products today. Some common examples:

•        Virtual assistants: Apple’s Siri, Amazon Alexa, and Google Assistant

•        Language models: AI tools such as ChatGPT, Claude and Gemini

•        Recommendation systems: whatever Netflix or Amazon decides to show you next

•        Computer vision: face unlock and apps that spot objects in photos

•        Robots: the Roomba, which maps a room and cleans it

•        Game players: Deep Blue, which beat Garry Kasparov at chess, and AlphaGo

These applications are part of the broader ecosystem of AI tools and software used for writing, search, recommendations, automation, and other everyday tasks. 

General AI (AGI)

General AI, sometimes called Strong AI, is the next step on paper. It would learn and reason about any subject the way a person does, without being retrained for each new job. It doesn’t exist. Not yet, anyway.

OpenAI, Google DeepMind, and Anthropic all name AGI as a long-term goal, but what they ship today is still Narrow AI, just a very flexible kind. Experts have different views on how close we are. Researchers like Michael I. Jordan have argued that spotting patterns at a huge scale isn’t the same thing as real reasoning.

The gaps are fairly specific. Current models can miss common sense in unfamiliar situations. They don’t set their own goals. And they don’t carry around a steady picture of the real world the way people do.

Superintelligence (ASI)

Superintelligence would beat humans at nearly everything at once: creativity, judgment, social skills, problem-solving. It’s purely theoretical. No prototype exists, and no test says any system is close.

People even argue about whether it’s possible. The idea still comes up constantly in AI safety and policy talks, because it’s easier to plan for the risks before something like it shows up.

Types of AI by Functionality

Reactive Machines

The simplest kind. No memory, no learning from the past. It reacts to whatever is in front of it right now. IBM’s Deep Blue judged each chess position on its own without recalling earlier games.

Limited Memory

This is where most modern AI lives. These systems use past data, usually through machine learning, to shape what they do next. A self-driving car is the classic case. It tracks the speed of nearby cars, the lane lines, and people crossing the street, then uses that short-term picture to make quick calls.

Large language models are often filed here too, since they learn from huge amounts of training data and can use your conversation as context. But go easy on the label. “Limited Memory” is a loose category, and it doesn’t mean the AI remembers things the way you do. Its knowledge comes from training, and it doesn’t reflect on its own experience. Systems also handle memory differently. So it’s fairer to say many current systems fit this box than to say all of them do.

Theory of Mind

Theory of mind is an idea from psychology. It means understanding that other people have their own beliefs, goals, and feelings. An AI with this skill would truly read someone’s mood and adjust, rather than spotting a keyword and firing back a canned line. It’s mostly research territory. Some emotion-detection tools and social robots take small steps toward it, but nothing available today can do it reliably.

Self-Aware AI

The last stage is a machine with its own consciousness and a sense of its own inner state. Nothing as it exists. It belongs more to philosophy than to any real engineering plan.

Modern Types of AI 

Terms like generative AI and agentic AI aren’t a third official framework. They describe what a system is used for, and they sit next to the two scales above, not in place of them. 

All of these are still Narrow AI.

AI Types vs. Technologies vs. Models

People mix these up all the time, so here’s a quick way to keep them apart. AI is the whole field. 

• Type is the big-picture label, like Narrow AI or AGI. 
• Technology is the method underneath, such as machine learning or deep learning. 
• A model is one specific trained system, like a particular language model. 
• An application is what people actually do with it, like a chatbot or a fraud alert.

You hear a few of these all the time, so let’s break them down: 

• Machine learning is when a computer learns patterns from data instead of following hand-written rules. 
• Deep learning is a kind of machine learning built on many-layered neural networks, and it powers most of today’s big breakthroughs. 
• Natural language processing (NLP) helps computers work with human language.
• Computer vision helps them make sense of images. 

None of these are “types of AI” in the official sense. They’re the tools that make the types possible.

How the Different Labels Fit Together

One AI system can wear several labels at once, because each label measures something different. Take ChatGPT. By capability, it’s Narrow AI. By functionality, it’s often described as Limited Memory. On the technology side, it’s built on deep learning and a transformer model. As an application, it’s generative AI, and in the way you use it, it’s conversational AI. Depending on the version, it can also be multimodal.

Once that clicks, the “too many types” problem mostly disappears. Nobody is contradicting anybody. They’re looking at the same thing from different sides.

Why ChatGPT Is Still Narrow AI?

This is probably the most common question on the topic, so let’s answer it straight. 

ChatGPT can do a lot. It writes, explains, and translates; some versions work with images, and some setups let it use tools. That range is impressive. But doing many things isn’t the same as being general intelligence.

AGI would mean learning almost anything a person can, adapting to truly new situations, and setting its own goals. ChatGPT still makes confident mistakes, struggles when a problem sits far from its training, and doesn’t pursue its own goals. Stanford’s 2026 AI Index points the same way, describing uneven progress: top models can win hard math contests and still trip over simple tasks like reading an analog clock. So it’s a very capable Narrow AI, and not AGI.

Which Types of AI Exist Today?

Why This Matters

Knowing the types helps you judge what a tool can really do. No matter what the marketing claims, every single AI tool a company can purchase today is still Narrow AI. So people still need to check the work, know where the data came from, watch for bias, and expect the occasional confident wrong answer. Treating a narrow tool like a wise expert is one of the most common and costly mistakes. Businesses evaluating AI adoption may also need help with selecting use cases, assessing risks, and planning implementation. AI consulting services can support that process.  If you want a practical starting point, NIST’s free AI Risk Management Framework is a good place to begin.

Conclusion

At its core, understanding the types of artificial intelligence comes down to two questions: how capable is the system, and how does it function? Narrow AI is the only capability-based category with widely deployed real-world systems today. Many modern AI applications can also be described using the Limited Memory category, although these functional labels are less precise for today’s complex systems. . General AI and superintelligence remain research goals, not products you can buy or use. Knowing the difference isn’t just a useless detail. It changes how you evaluate every AI tool you come across, from a chatbot at work to a self-driving car on the road.

Frequently Asked Questions

Sources

1.       IBM: Types of Artificial Intelligence – https://www.ibm.com/think/topics/artificial-intelligence-types

2.       Stanford HAI: The 2026 AI Index Report – https://hai.stanford.edu/ai-index/2026-ai-index-report

3.       NIST: AI Risk Management Framework – https://www.nist.gov/itl/ai-risk-management-framework