Types of AI in 2026: narrow, general and super AI explained

Here is the short version. There are three main types of AI by capability: narrow AI, which does one job or a small set of jobs; general AI (AGI), which would match people across almost any task; and superintelligence, which would beat them. Only narrow AI exists today. Everything you use, from spam filters to chatbots like ChatGPT, falls in that first bucket, even if the newest models are starting to blur the line.

The confusion comes from the fact that people sort AI in more than one way. You will see "3 types of AI", "4 types of AI" and "7 types of AI" lists, and they are often counting different things. This guide gives you a simple mental model for all of them, real examples for each, where each type breaks, and what it means for the tools you pick. It is based on published frameworks and reports, checked in late October 2026. If you want the basics first, start with our explainer on what AI is.

Short answer

The different types of AI are usually grouped in two ways:

• By capability (how broad it is): narrow AI, general AI and super AI. IBM describes narrow AI as "the only type of AI that exists today."
• By functionality (how it handles memory and the world): reactive machines, limited memory, theory of mind and self-aware AI. This four-part scheme comes from computer scientist Arend Hintze, writing for the World Economic Forum in 2016. Only the first two exist.

Put the two lists together and you get the "7 types of AI" you see in many articles. Generative AI, the kind that writes text and makes images, is not a separate type on either axis. It is narrow, limited-memory AI.

How the types of AI work: three ways to sort them

Think of it as three different rulers. Each one measures a different property of the same system, so one tool gets a reading on all three.

Ruler 1: Capability (narrow, general, super)

Narrow AI, also called weak AI, is trained for a single task or a small family of tasks. IBM notes it can perform that task "often far faster and better than a human mind can," but "it can't perform outside of its defined task." A chess engine cannot file your taxes. A fraud model cannot read an X-ray.

General AI, or strong AI, would take what it learned in one area and apply it to new problems, the way you can learn to drive and then pick up a forklift with little extra training. It does not exist yet.

Super AI, or artificial superintelligence, would beat the best humans at nearly everything. It is purely theoretical.

Ruler 2: Functionality (the four Hintze types)

Hintze's four types sort systems by how they use memory and whether they model other minds:

1. Reactive machines have no memory. They look at the current input and respond. IBM's Deep Blue, which beat Garry Kasparov at chess in 1997, is the classic case: it read the board and picked a move without learning from past games.
2. Limited memory systems use recent data for a short time. Hintze's example is a self-driving car tracking the speed and direction of nearby vehicles before changing lanes. The car does not keep that as lifelong experience.
3. Theory of mind AI would understand that people have beliefs, emotions and intentions. Research exists, but no system does this in a robust way.
4. Self-aware AI would have a model of itself and its own internal states. Hintze wrote that researchers are probably far from building it.

Ruler 3: Technique (how it is built)

Many lists also count methods as types: machine learning, deep learning, natural language processing, computer vision, robotics and expert systems. These are not levels. They are engineering approaches, and one product often stacks several of them. A robotaxi uses computer vision to see, deep learning to predict, and classic planning code to steer. For the difference between the first three terms, see our guide to AI vs machine learning vs deep learning.

Here is how the rulers line up:

A worked example: classify a chatbot

Take a modern assistant and run it through each ruler.

• Capability: It handles writing, coding, summarising and maths, which feels general. But it cannot drive a car, fold laundry or reliably learn a new skill from a single attempt. IBM's position is that ChatGPT "is considered a form of Narrow AI." Google DeepMind researchers took a different view in a 2023 paper: their "Levels of AGI" framework placed ChatGPT and Bard at the lowest rung, "emerging AGI". Both are defensible. They just draw the line in different places.
• Functionality: It reads the conversation in front of it (its context window) and responds. When the chat ends, that working memory is gone unless a separate memory feature saves notes. That is limited memory, Type II.
• Technique: Deep learning, specifically a large language model trained on text, plus natural language processing. Our explainer on what an AI model is covers how that training works.

Result: narrow (or at most "emerging general"), limited memory, deep learning. That three-part label tells you far more than "it's AI".

Real examples

Narrow AI is everywhere, and much of it is invisible.

• Recommendation engines. IBM lists Netflix's recommender as reactive AI: it scores what to show you from the data available now.
• Writing tools. Grammar checkers spot errors and suggest rewrites in one narrow domain. Our roundup of grammar AI tools shows how specialised they are.
• Science. DeepMind's AlphaFold predicts protein structures. The same DeepMind framework rates it as narrow AI at the "superhuman" level: better than any person, but only at that one job.
• Self-driving cars. Waymo was giving about 400,000 paid rides per week across six cities in February 2026, with a goal of more than 1 million weekly rides by the end of the year. Each ride is limited-memory AI tracking objects for seconds at a time.
• Chat assistants. OpenAI began rolling out GPT-6 to ChatGPT users on 7 October 2026, with a "Sol" model for paid plans and "Luna" for Free and Go users. It is the most general-feeling AI most people will touch this year, and it is still classed as narrow by most definitions.
• Agents. Systems that plan, call tools and act over several steps are often sold as a new "type". On the capability ruler they are still narrow; what changes is that they act, not just answer. Our piece on agentic AI explains the loop.

Benefits and limitations

The benefit of narrow AI is focus. Trained on one job, a system can beat people at it on speed, cost or accuracy. The 2026 Stanford AI Index, as reported by R&D World, found top models reached 93% on GPQA, a set of graduate-level science questions, against an expert human baseline of 81.2%. Some health systems reported up to 83% less time spent writing clinical notes.

Where it breaks

The same report describes a "jagged frontier": strong peaks next to surprising valleys. A few numbers make the point.

Sources: R&D World for GPQA and ClockBench; Stanford HAI for OSWorld, where AI rose from about 5% in 2024 to 66.3% in 2025.

Now do the maths an engineer would. A 66% success rate per task sounds close to human. But if an agent has to chain five such steps, and each succeeds independently 90% of the time, the whole job succeeds only 0.9 × 0.9 × 0.9 × 0.9 × 0.9 ≈ 59% of the time. Small error rates compound fast. That is the gap between "impressive demo" and "safe to leave running".

Other common failure modes:

• Out-of-scope inputs. Narrow systems fail quietly when the input is unlike their training data, such as a rare medical image or a road layout the car has not seen.
• Confident errors. Language models can produce fluent, wrong answers. Fluency is not evidence of accuracy.
• No lasting learning. Limited-memory systems forget. If a tool seems to "remember" you, a separate memory feature is storing notes, and you should know where they live.
• Real-world harm. Stanford HAI reports that recorded AI incidents reached a record 362 in 2025.

What it means for you

Knowing the types of AI is useful mainly because it sets your expectations. A quick checklist before you trust a tool:

1. Name its job. If you cannot state the narrow task it was built for, you cannot judge whether it is good at it.
2. Check its memory. Does it keep data between sessions? Where, and can you delete it?
3. Test it on your edge cases. A tool that scores well on benchmarks may still fail on your odd invoices or regional accents.
4. Keep a human check where errors are costly. Money, health, legal and safety decisions need review, whatever the marketing says.
5. Ignore "AGI" in sales copy. There is no agreed test for it, so a vendor claim tells you little.

And a small decision guide for the label you will see most:

Regulators sort AI differently again: by risk, not capability. The EU AI Act bans some practices outright and adds duties for "high-risk" uses. After a 2026 revision, the high-risk deadlines moved to 2 December 2027 for stand-alone systems and 2 August 2028 for AI built into regulated products, according to law firm Gibson Dunn, while transparency rules start from 2 August 2026. Our AI regulation guide covers the wider picture.

What happens next

The real argument is whether, and when, narrow AI turns into general AI. Forecasts have shortened sharply. A review by 80,000 Hours notes that a 2023 survey of thousands of AI researchers gave a median 50% chance of "high-level machine intelligence" by 2047, while Metaculus forecasters, as of February 2026, put a 50% chance of AGI at 2033. The same review points out that AI company leaders talk of two to five years, and that they have an incentive to be optimistic.

Treat these as opinions, not schedules. The definitions differ, and the jagged frontier means a system can be superhuman at science quizzes and still misread a clock. The more likely near-term change is not a new "type" but broader narrow AI: models that handle text, images, audio and actions in one system, with longer memory. Theory-of-mind and self-aware AI remain research questions with no working examples.

For how we got here, from Deep Blue to today's chatbots, see our history of who created AI technology.

FAQ

What are the different types of AI?

By capability there are three: narrow AI, general AI and super AI. By functionality there are four: reactive machines, limited memory, theory of mind and self-aware AI. Only narrow AI, and within it reactive and limited-memory systems, exists today. Lists of "types" by technique, such as machine learning or computer vision, describe how AI is built rather than how capable it is.

What is narrow AI?

Narrow AI, or weak AI, is a system trained to do one task or a small set of related tasks, such as filtering spam, recommending films, predicting protein shapes or writing text. It can outperform people at that task but cannot transfer its skill to unrelated jobs. All AI products available today are narrow AI by most definitions.

What types of AI are there?

In practice you meet predictive AI (scores, ranks and flags), generative AI (makes text, images, audio and video), conversational AI (chatbots and voice assistants), computer vision (understands images), robotics and AI agents (takes multi-step actions). All of these are narrow AI. General and super AI remain hypothetical.

How many types of AI are there?

It depends on the scheme. The common answer is seven: three capability types (narrow, general, super) plus four functionality types (reactive, limited memory, theory of mind, self-aware). They overlap, because they measure different things. Of the seven, only three describe systems that exist: narrow, reactive and limited memory.

What type of AI is generative AI?

Generative AI is narrow AI with limited memory, built with deep learning. It creates new content from patterns in its training data. IBM notes that tools such as ChatGPT rely on limited-memory capabilities. Some researchers, including a Google DeepMind team, argue the most capable chatbots count as "emerging" general AI, but this is a minority view.