Two Different Breeds, Two Different Logics

In the previous part we said that the thing you chat with and the thing that detects a lesion on a radiograph belong to two different generations and arrive at their answers by two different logics. Now let’s get to what those two logics actually are, because everything that follows in this book grows out of this one difference.

First, let’s use the right name. The thing you chat with — ChatGPT, Google’s Gemini, Claude — is a kind of artificial intelligence called a large language model, or LLM for short. From here on, whenever we say “the model” or “LLM,” we mean this family; the same thing you already know, just under its technical name.

Now the difference. Picture that diagnostic software meant to find caries on a radiograph. How was it built? It was shown thousands, sometimes hundreds of thousands, of radiographs on which specialists had marked where caries was and where it was not. The model guessed again and again, and each time its guess was compared against the specialist’s correct answer: right or wrong? That cycle was repeated until the model became accurate at detection.

Here is the key point: a correct answer exists, outside the model. The caries either really is there or it is not, and that is settled by the reality of the patient’s tooth, not by the model. The entire training of this system means moving closer to that external truth.

Now the LLM. This one was not trained that way at all. It was given an enormous volume of text, from books and papers to web pages and conversations, and its job throughout training was one thing only: guessing the next word. A sentence was placed in front of it half finished and it had to guess what the next word should be. Billions of times. “The sky is ___” and it learns that “blue” is more probable than “not blue.” That is all, at an astronomical scale.

And this difference is not a simple technical one. It is two completely different goals. That diagnostic software spent its whole life on one question: what is correct? The language model spent its whole life on a different question: what is probable to come next?

Note that these two often arrive at the same answer, and that is exactly what makes the language model so useful. Because what is correct has usually been repeated a great deal in the texts the model learned from, and is therefore also “probable.” When you ask what the capital of France is, the correct answer and the probable answer are the same, because “Paris” has appeared next to that question thousands of times. That is why the model gives a correct answer most of the time.

But this overlap is not a guarantee, it is a stroke of luck. Where “probable” and “correct” come apart — and they do come apart — the model immediately takes the side of “probable.” Because that is fundamentally what it was built for. Correctness was never its goal.

That is why you cannot treat an LLM the way you treat that diagnostic software. That one is a tool tuned to reach a correct answer. This one is a tool tuned to produce plausible text, and being plausible is not the same as being correct.

And precisely here a question arises: if the model is not after correctness and only after probability, then what does it do when it does not know something? It does not stay silent. The answer to that question takes us to one of the most important things a dentist should know about this tool, which is the subject of the next part.