Not always the likeliest

At the bottom of its range the dial makes the model take its top-scoring token every time, so the same question gets the same answer forever. Turn it up and the model itself is untouched. Same weights, same scores, nothing added and nothing taught. The only change is which token wins the draw.

The second choice can win. So can the third, or the fourth. The higher the dial goes, the further down the list it will reach, until it is picking tokens that were never plausible and the output stops making sense.

So there is a range in the middle where the model is neither repeating itself predictably nor babbling gibberish, and that range is where anything you would call good writing comes from.

In practice the dial sits in that middle, and the tail is usually cut off as well, so the tokens actually in play are a handful of reasonable continuations. The choice is rarely between sensible and absurd. It is between several words that would each have worked.

And nothing in there is deciding which of those words would be better. The dial moves one number, the likelihoods shuffle, and a draw is taken. If a model picks a word that strikes you as inspired, no part of the machine picked it for being inspired.

Try asking the same question twice with the dial at its lowest setting. You get the same answer both times. Ask again with the dial turned up a little, and you usually get a different answer.