Let me start with a game, and I promise it goes somewhere. Sing this line in your head: "The wheels on the bus go..."

If your brain just finished that with "round and round," then congratulations, because you have just done the single thing that sits underneath every large language model on the planet. You predicted the next most probable word. You didn't reason about buses, you didn't consult your feelings about public transport, you simply reached for the word that almost always follows, and you were right.

That is the whole trick, and it is worth understanding properly, because the gap between what AI actually does and what people imagine it does is where most of the bad decisions about it get made.

The idea underneath everything

At its simplest, this kind of prediction is what mathematicians call a Markov chain, which is a rather grand name for a very ordinary idea. A Markov chain looks at the state you are in right now and works out what is most likely to come next, based purely on probability. Weather models use them, board games use them, and the earliest text-prediction tools used them too. Given "the wheels on the bus go," the odds of "round and round" are enormous, so that is what gets picked.

If that were the end of the story, AI would be a party trick and nothing more. A pure Markov chain has almost no memory, because it only really cares about the state it is in this instant, and that is exactly why the autocomplete on your phone a decade ago was so easy to laugh at.

Where modern AI pulls ahead

Modern AI does not just react to the last word it saw, and that single difference is what separates it from its ancestors. It breaks your sentence into small pieces (the jargon is "tokens," though you can think of them as chunks), and before it writes a single word back it reads the whole of what you gave it. Then it builds its answer one piece at a time, and at every step it weighs the entire context around those chunks rather than just the word immediately before. This whole process of running the model to turn your question into a response is what engineers call inference.

So when it sees "the wheels on the bus go," it is not only noticing the word "go." It is picking up that you are reciting a specific nursery rhyme, that nursery rhymes tend to repeat, that this one has a rhythm, and that millions of examples of it exist in everything it has read. All of that context is what lets it land on "round and round" with such confidence, instead of guessing something absurd like "the wheels on the bus run out of specification after forty thousand miles."

The Markov chain is really the ancestor. What we now call AI is what happens when you take that simple next-word guess, give it an enormous memory, and let it read the whole room before it answers. Scale the idea up far enough and it stops looking like autocomplete and starts to look, a little unnervingly, like understanding.

It is not magic, and that is the point

I labour this because the mystique is expensive. When people believe AI is thinking, they tend to do one of two unhelpful things. They over-trust it, handing it decisions it has no business making, or they panic about it in ways that don't lead anywhere useful either. When you understand that it is really probability scaled up until it resembles understanding, you get calmer and more effective at the same time.

You learn to check its confident-sounding answers, because confidence is precisely what it is built to manufacture, whether or not it happens to be right. You learn that it will be brilliant at the predictable and shaky at the genuinely novel, which tells you where to lean on it and where not to. And you learn that the context you give it changes the quality of what you get back, because context is the whole engine, so a vague question gets a vague guess and a well-framed one gets something far closer to what you needed.

So the next time a chatbot writes something that makes the hair on the back of your neck stand up, remember the bus. It isn't thinking. It is doing, at a scale we can barely picture, exactly what you did when you finished that line without a second thought. The question I keep coming back to, and I don't think it is a comfortable one, is how much of our own fluent, confident, everyday thinking is really doing anything more than that.