What are the main types of AI?

“Types of AI” can be confusing because people use the phrase to mean different things. There are really two standard ways experts categorise AI — by capability and by function — plus a set of techniques that often get muddled in. Let’s untangle all three.

How is AI categorised by capability?

The capability framework asks: how flexible and powerful is the intelligence? It has three levels:

• Narrow AI (weak AI) — built for one specific task. This is the only kind that exists today, and it covers everything from spam filters to chatbots. See narrow AI vs general AI for a deeper look.
• General AI (AGI) — would match human flexibility across any task. Hypothetical; not yet built.
• Superintelligence — would surpass the best humans at virtually everything. Purely theoretical.

The key takeaway: despite the hype, all real-world AI is narrow AI.

How is AI categorised by function?

The function framework, often called the “four types of AI”, asks: how does the system handle information and experience?

Reactive machines

The simplest type. They respond to the current situation with no memory of the past. A classic example is a chess computer that evaluates the board in front of it but doesn’t learn from previous games.

Limited-memory AI

These systems learn from recent data to make better decisions. This describes almost all useful AI today — self-driving features, recommendation engines, and chatbots all draw on past examples through machine learning. Their “memory”, though, is limited and task-specific.

Theory-of-mind AI

A hypothetical future type that would understand that people have emotions, beliefs, and intentions, and adjust its behaviour accordingly. Genuine versions don’t exist yet.

Self-aware AI

The most speculative type: AI with consciousness and a sense of self. This belongs to science fiction and philosophy, not current engineering.

Only the first two — reactive and limited-memory — actually exist.

Where do machine learning and deep learning fit?

This is where many explanations get tangled. Machine learning, deep learning, and generative AI are not separate “types” in the frameworks above — they’re techniques and capabilities:

• Machine learning is the main method used to build today’s AI.
• Deep learning is a powerful sub-method using large neural networks.
• Generative AI describes what a system does — create new content like text or images.

So a single tool can be described in several ways at once. ChatGPT is narrow AI (capability), a limited-memory system (function), built with deep learning (technique), used for generative tasks (function/output). All four labels are correct and describe different aspects.

A simple way to remember it

When someone mentions a “type of AI”, ask which lens they mean:

• How smart/flexible is it? → capability (narrow / general / super).
• How does it use information? → function (reactive / limited-memory / theory-of-mind / self-aware).
• How is it built or what does it do? → technique (machine learning, deep learning, generative AI).

Keeping these three lenses separate clears up most of the confusion around AI categories.

Why are there so many overlapping labels?

The confusion is understandable. AI is studied by computer scientists, philosophers, and businesses, and each group brings its own vocabulary. The capability framework comes largely from debates about how powerful AI could become. The function framework is a popular way to describe how systems handle information. The “technique” terms come from engineering.

Because these schemes describe different aspects of the same systems, they overlap rather than compete. That’s why a single product can wear several labels at once without contradiction.

Which type should you care about?

For most people, the practical distinctions are simple:

• Narrow vs general tells you whether something is real today (narrow) or hypothetical (general). This matters most for cutting through hype.
• The technique (machine learning, deep learning, generative AI) tells you roughly how a tool was built and what it does.

The function categories are useful background, but in everyday life the narrow-vs-general split and the technique labels are the ones that help you understand what a given AI tool actually is — and isn’t.

In one sentence

AI is sorted by capability (narrow, general, super) and by function (reactive, limited-memory, and two hypothetical types), while machine learning, deep learning, and generative AI describe how systems are built — and today, everything real is narrow, limited-memory AI.