Definition
A common teaching framework, popularised by Arend Hintze, sorts AI into four types by how it functions: reactive machines, limited memory, theory of mind and self-aware AI. The first two exist, and the last two are still ideas.
In simple words
A reactive machine only sees the present, like a thermostat with no memory. Limited memory uses recent data, like a self-driving car watching nearby cars. Theory of mind would understand what people feel and want. Self-aware AI would know it exists. We've built the first two.
Why it matters
This list appears in many syllabi, exams and interviews. It's also useful for spotting hype: when a chatbot appears to have feelings, that's pattern-based text, and nothing shows it's self-aware.
Key terms
- Reactive machine
- An AI system that responds only to what is in front of it now, with no memory of the past.
- Limited memory AI
- An AI system that uses past data, either from its training or from the last few moments, to decide better.
- Theory of mind
- The ability to understand that other beings have their own beliefs, wishes and feelings. No AI has it yet.
- Self-aware AI
- An imagined AI that would have consciousness and a sense of itself. It does not exist.
- Consciousness
- The experience of being aware: of feeling, sensing and knowing that you exist.
Explained in detail
A second way to sort AI
The previous topic sorted AI by capability: how much it can do (narrow, general, super). This topic sorts AI by functionality: how it works inside, especially how it uses memory and how it models the world. The four-type list was popularised in 2016 by Arend Hintze, a researcher who studies artificial intelligence and evolution. The two ways of sorting answer different questions, and one system can be placed in both. A chess computer, for example, is narrow by capability and reactive by functionality.
Type 1: reactive machines
A reactive machine looks at the present situation and chooses a response. It keeps no record of what happened before and cannot use past experience. IBM's Deep Blue, which beat world chess champion Garry Kasparov in 1997, is the classic example. It examined the board in front of it and searched through millions of possible moves each second, but it did not remember earlier games or learn from them. Reactive machines are reliable, because the same situation always gives the same response, but they cannot improve by themselves.
Type 2: limited memory
Almost every AI system in use today has limited memory. It learned from past data during training, and many also keep track of recent events for a short time. A self-driving car watches how fast nearby vehicles have been moving over the last few seconds in order to predict where they will be next. A chatbot keeps the earlier messages of your conversation so that its reply fits. The memory is called limited because it is temporary or fixed: the car does not remember yesterday's journey, and most chatbots do not permanently learn from each chat.
Type 3: theory of mind
Theory of mind is a term from psychology. Children develop it at around four years of age, when they begin to understand that another person can believe something different from what they themselves know. An AI with theory of mind would truly understand that people have beliefs, intentions and emotions, and would adjust its behaviour to them, as a good teacher notices that a learner is confused before the learner says anything. Today's chatbots can produce sentences about feelings, and some pass written tests designed for this skill, but researchers disagree about whether this shows real understanding or learned patterns of words. No system is established as having it.
Type 4: self-aware AI
The last type is AI that would be conscious: aware of itself, with its own inner experience. This exists only in stories and thought experiments. Scientists do not yet agree on what consciousness is or how it arises in the human brain, so nobody knows how to build it, or even how to test for it reliably. When a chatbot writes "I feel happy", it is producing words that fit the conversation; this is not evidence of a feeling.
How it works
| Type | How it works | Status | Example |
|---|---|---|---|
| Reactive | Responds only to the present input and stores no experience | Exists | IBM's Deep Blue chess computer |
| Limited memory | Uses past data, for training or for a short time | Exists: nearly all modern AI | Self-driving cars, chatbots |
| Theory of mind | Would model other minds' beliefs and emotions | Research only | None established |
| Self-aware | Would have consciousness and a sense of self | Hypothetical | None |
Try it yourself
How much can it remember and understand?
This is another way to group AI. Only the first two exist today.
Which type is it?
Sort each system by how it works with memory and minds.
A video app recommending based on what you watched
A self-driving car watching nearby cars
IBM's Deep Blue chess computer (1997)
A machine conscious of its own existence
A robot that truly understands your feelings and intentions
Common mistakes
- Thinking that a chatbot which says it has feelings really has them.
- Mixing up this list with narrow, general and super. Those measure capability, and this measures function.
Check yourself
Summary
- Four types: reactive, limited memory, theory of mind, self-aware.
- Nearly all of today's AI is limited memory.
- It's a teaching framework, separate from narrow, general and super.
Finished reading?