The 5 main types of artificial intelligence are reactive machines, limited memory, theory of mind, self-aware AI and artificial general intelligence. Only reactive machines and limited memory systems exist today; theory of mind, self-aware AI and AGI remain experimental or theoretical.
This guide explains why the ladder is built that way, what each rung looks like in a machine, and how the types map onto robots. For the engineering side, see the guide that asks how is ai used in robotics.
Two ways to sort AI
The five-type list ranks systems by how they use memory and how much of the world, including other minds, they can model. A second scheme ranks them by overall capability. EDC Paris describes a capability-based approach with three categories: ANI, AGI and ASI. It adds that only the first of those exists today.
The two schemes overlap rather than compete. QA defines narrow AI as systems that perform specific tasks without the ability to learn beyond their intended purpose. Reactive machines and limited memory systems both fall inside that narrow band. Teachfloor makes the same point with a robotics-adjacent case: self-driving cars, while sophisticated, count as narrow AI because they are designed for one job, driving. It also calls narrow AI the most common type today.
The practical lesson is that a system can be impressive and still sit low on the ladder. When you judge a product or a robot, ask what it remembers and what it can model, not how advanced it sounds.
The types that exist today
Reactive machines
Reactive machines are the simplest rung. EDC Paris notes that they have no memory and cannot learn from past experiences. The standard example is IBM’s Deep Blue, the chess computer that defeated world champion Garry Kasparov in 1997. It could not improve its performance beyond the rules programmed into the system.
Dilitrust adds the limit on scope: Deep Blue could not perform any task beyond chess. Winning a world-championship match and being unable to do anything else is the defining trade-off of this type.
For robots, the implication is predictability. A machine whose behaviour comes only from programmed rules acts the same way every time, which suits tightly defined jobs. The cost is that it cannot adapt when conditions change, so any robot built this way needs its designers to anticipate the situations it will face.
Limited memory systems
Limited memory is where most working AI sits. Dilitrust describes these systems as ones that learn from past information to improve future responses. EDC Paris lists the main examples: self-driving cars, fraud detection systems and generative AI tools such as ChatGPT all rely on limited memory.
The self-driving car is the clearest robot case. QA explains that such a car observes other vehicles so it can navigate the road and adjust accordingly. A reactive controller would respond only to the instant in front of it. A limited memory system carries recent observations forward and uses them to adjust.
That shared category covers very different machines, and the label does not tell you how good any of them is. A fraud detector and a chatbot are both limited memory; the mechanics of the second are covered in Generative AI: How It Works — A Practical Guide To Models, Training, And Applications (2026).
The types still ahead
Theory of mind
Dilitrust defines theory of mind AI as a system that would understand that entities in the world have their own beliefs, desires and intentions. QA gives a driving illustration: a person who knows a child plays near a neighbour’s driveway slows down when passing it, and a theory-of-mind car would be expected to do the same. The car would be reasoning about a person, not just detecting an object.
This is the difficult part. QA observes that emotional cues are highly complex and that AI may take a significant amount of time to master them. EDC Paris states that this type of AI has not yet been achieved and remains largely experimental. A product claiming to truly understand human intent is describing an aim, not a delivered capability.
Self-aware AI
EDC Paris says self-aware AI exists only in theory and is often discussed in relation to ethics. It has no working example and no robot that embodies it. Treat it as a topic for ethical debate, not as something to plan a project around.
Artificial general intelligence
AGI is the type people ask about most. EDC Paris defines it as a future generation of AI capable of reasoning, learning and adapting like a human being, and states that it remains a research objective rather than a commercial reality. Teachfloor puts the benchmark simply: AGI would be able to perform any intellectual task that a human can.
QA describes general AI as human-like intelligence that can execute multiple tasks simultaneously, and adds that we are still considerably distant from constructing an AGI system. Dilitrust frames it as the ability to apply knowledge across different domains and treats it as theoretical. Compare that with Deep Blue: a chess champion that could do nothing else is the opposite of general intelligence.
The learning methods behind the types
Reactive, limited memory and the other types describe what a system can do. Machine learning describes how a system gets there, so these are training methods rather than extra rungs. testRigor defines machine learning as the science of allowing computers to learn without being programmed to do so. That is the reason limited memory systems can improve while Deep Blue could not.
- Supervised learning. The model trains on labelled datasets in which the inputs and outputs are known, so it learns a mapping it can apply to new inputs.
- Deep learning. This is a class of machine learning frameworks that use multi-layered neural networks to model complex patterns in data.
- Reinforcement learning. The system learns by maximising a score. testRigor’s examples are teaching robots to walk and teaching AI agents to play chess, which makes this the method most closely tied to robot behaviour.
Chess shows up on both sides of the divide. Deep Blue followed programmed rules, while a reinforcement-learning chess agent improves by chasing a score. For a related look at trained models judging text, see How Does AI Detection Work? A Clear, Practical Guide For 2026.
The types compared, with robot examples
The table maps each type to what it relies on, where it stands and what that means when you evaluate a robot or an AI tool.
| Type | Defining trait | Status today | Machine example | Planning use |
|---|---|---|---|---|
| Reactive machines | No memory; cannot learn from past experiences | Exists | Deep Blue, the chess computer that beat Kasparov in 1997 | Suits fixed, well-defined tasks where predictable rule-following matters more than adaptation |
| Limited memory | Relies on limited memory of past information | Exists | Self-driving cars, fraud detection, ChatGPT-type tools | The default for any robot or tool that must adjust to changing conditions |
| Theory of mind | Would model other people’s beliefs and intentions | Not yet achieved; largely experimental | Illustrative only: a car slowing near a child playing by the street | Do not treat as a purchasable feature |
| Self-aware AI | Would have awareness of itself | Exists only in theory | None | Relevant to ethics discussion, not to project planning |
| AGI | Would reason, learn and adapt like a human across tasks | Research objective, not commercial | None built | Do not plan around it |
Table entries draw on EDC Paris and QA.
Read the table as a decision aid. If a task has fixed rules and little variation, reactive behaviour may be enough, and it is the easiest to predict. If conditions shift, you need a limited memory system that learns from past data. Anything above that is a research direction, so a vendor selling it as a finished product deserves scrutiny.
Frequently asked questions
What are the types of AI?
They are reactive machines, limited memory, theory of mind, self-aware AI and artificial general intelligence. They run from systems with no memory to systems that would model other minds, themselves, and any intellectual task.
Which types of AI exist today?
Reactive machines and limited memory systems. Theory of mind is largely experimental, self-aware AI exists only in theory and AGI is still a research objective.
What is AGI?
AGI, or artificial general intelligence, is a future form of AI that could reason, learn and adapt like a human being across different domains. It has not been built, and sources describe it as considerably distant.
Is ChatGPT an example of AGI?
No. EDC Paris lists ChatGPT-type generative tools among limited memory systems, while AGI remains a research goal rather than a commercial product.
