What Comes Next
Finish this sentence: the cat sat on the…
You said “mat”. Of course you did. You’ve read that sentence a thousand times, and your brain has quietly worked out that “mat” is far more likely than “sofa”, and “sofa” is far more likely than “moon”. You didn’t decide anything. You predicted.
That small act sits at the heart of the AI everyone is talking about. A large language model has been trained on an enormous amount of human writing, and what it learned to do is predict what comes next. Give it some words and it works out the likely next word, then the one after that, and so on, at a speed and scale that would make your phone’s predictive text weep with envy.
Language is surprisingly good at being predicted. In 1951, the mathematician Claude Shannon ran experiments asking people to guess the next letter of English text, and found they could do it far better than chance. English is full of patterns: grammar, habits, clichés and the phrases that turn up in every second strategy document. Once you’ve seen enough of it, a great deal of what people write turns out to be guessable.
That’s also why AI can feel so uncanny. Predicting language well at scale lets it summarise a report, draft a reasonable email and explain an idea in plain English. Whether you want to call that intelligence is a separate argument, and a mostly unproductive one. The more useful question is what follows from knowing it’s built on prediction.
The first thing is that likely isn’t the same as true. When an AI tells you a fact, it’s giving you the most plausible-sounding answer, not one it has checked. Most of the time the plausible answer and the true one are the same, because true things get written down a lot. When they aren’t, the answer sounds exactly as confident either way. It’s a bit like that colleague who’s always certain and right about seventy per cent of the time. You’d still listen to them, but you’d check before you put it in the board paper.
The second is that prediction drifts towards the middle. Ask a vague question and you’ll get the most likely response to a vague question, which is the average of everything ever written on the subject. If your marketing plan reads like every other marketing plan, that’s not the machine being lazy. It’s the machine being accurate about what you asked for.
The third is the good news. Because the answer depends on the words that come before it, your words change what’s likely. Tell it who the audience is, what you actually mean, what shape you want and what to avoid, and you shift the odds towards something specific and useful. That’s not a technical skill. It’s the same thing good managers do when they brief a person, and the same thing good writers do with a reader.
So when someone describes AI as magic, or as a menace, I find it helps to come back to the cat on the mat. It’s very, very good at guessing what comes next. That makes it remarkably useful, as long as you remember that the next likely word and the right one aren’t always the same, and that the best way to improve the guess is to say more clearly what you want.
