Reef TRHOn 11 May 1997, an IBM machine called Deep Blue did something no computer had managed before: it beat the reigning world chess champion, Garry Kasparov, over a full six-game match. The papers called it thinking. It wasn't. Deep Blue had no idea it was playing chess, no memory of the game before, no sense that Kasparov was even in the room. It weighed roughly 200 million board positions a second and returned the strongest move. That gap, between a machine that looks intelligent and one that understands anything at all, is the whole story of AI, and it is why researchers tend to sort the field into four types.
It helps to read those four as a ladder rather than a menu. The bottom two rungs are real, commercial, and almost certainly already running somewhere in your business. The top two are a mix of open research and outright science fiction. Knowing which rung a system stands on tells you far more about what it can actually do than any product page ever will.
1. Reactive machines
The bottom rung is pure reflex. A reactive machine takes an input and returns an output by a fixed set of rules, with no memory of the past and nothing learned from it. Deep Blue is the textbook case: extraordinary at chess, and completely incapable of doing, or even knowing, anything else.
Systems this pure are rare on their own now, but the pattern, the same input producing the same output every time, still lives inside larger tools. A spam filter scoring a single email. The rules engine behind a checkout. A thermostat deciding to switch on. They are fast, dependable, and constitutionally unable to surprise you, which is a feature, not a flaw, when the job is well defined.
2. Limited memory
Climb one rung and machines begin to learn from history. A limited-memory system is trained on past data and uses it to act on situations it has never met. Waymo's self-driving cars, which have now logged tens of millions of driverless miles, read the road by drawing on everything they have seen before. So does the recommendation engine that queues your next programme, and the model that flags a fraudulent transaction a half-second after you tap.
Here is the part the marketing tends to skip: this rung is where almost every useful AI actually stands, including the large language models behind today's chatbots, GPT-4 and Claude among them. They learned patterns from a fixed body of past text, and a chatbot's "memory" of your conversation lasts only as long as its context window. That is not a knock; rung two is the most commercially valuable rung on the ladder by a wide margin. But it does mean a system is only ever as good as the data it learned from, which is why, in practice, most AI projects live or die on data quality rather than model choice.
3. Theory of mind
The third rung is where things stop being shipped and start being argued about. A theory-of-mind AI would grasp that other people hold their own beliefs, intentions and feelings, and adjust to them in real time. Today's tools can imitate the surface of this, an emotion-recognition model can read a tense face or a clipped tone, but that is pattern-matching, not a real model of what another mind is thinking. Whether the largest language models show anything like genuine theory of mind is an active and contested research question, and the honest answer for now is: not yet. If it ever arrives, the systems that negotiate, coach, or handle a delicate support call will change beyond recognition.
4. Self-aware
The top rung is the one the films care about and engineers lose no sleep over. A self-aware AI would have a sense of itself: consciousness, an inner life, feelings of its own. There is no known engineering path to it, and the live questions are still mostly for philosophers. For now it belongs on the poster, not the roadmap.
Where today's tools actually sit
Line them up and the picture is clear. Rung one is a dependable reflex. Rung two learns from the past and quietly runs most of the AI you can actually buy in 2026. Rung three is a research frontier. Rung four is fiction. The confusion, and the expense, arrives when a rung-two product gets described in rung-four language: "our AI understands you", "it thinks like your best analyst". The distance between what a system does and what it sounds like is exactly where budgets tend to disappear.
Why the rung matters when you're buying
Place a tool on the ladder and the right questions change. If it is limited-memory, and it almost always is, the useful questions have nothing to do with consciousness and everything to do with data: what was it trained on, how closely does that resemble your reality, and what happens the first time it meets an example it has never seen? Match the rung to the problem and AI becomes one of the sharpest instruments you can hand a business. Mistake rung two for rung four and you will buy a demo that never survives contact with real work.
So the next time someone tells you their system thinks, or understands, or feels, you have a simple test. Ask what it remembers, and what it learned from. The answer usually places it one rung above Deep Blue, which is a great deal further than it sounds, and still a very long way from the robot on the poster.