It's a Tool, Not a Guide

So far we have seen one thing: the software that reads a radiograph and the model you chat with are two different kinds of things, with two different logics, and two different kinds of error. And we saw that the model can, with complete confidence, build something that doesn't exist. Now the real question is: how should you work with something like this?

The answer is one word: like a tool, not like a guide.

The difference between the two is subtle but crucial. You trust a guide to know the way and take you there; you yourself don't need to know where you're going. But a tool only works in the hands of someone who already knows what they're doing. A dental handpiece in your hands, as a dentist, is a tool; in the hands of someone who doesn't know where to cut, it's a dangerous thing. The tool itself hasn't changed; what changes is the awareness of the person holding it.

It's exactly the same with the language model. Remember we said that the only thing that gives hallucination away is your own knowledge? That means the model is only a useful tool for you within a domain you yourself are already competent in. If you ask it about something you aren't yourself fluent in, you no longer have any way to tell whether its answer is correct or fabricated. At that point it's no longer your tool; you've become its captive. The same model is a powerful tool in one domain and a dangerous trap in another, and the boundary between the two is your own awareness.

What's interesting is that this isn't just a personal opinion. The American Dental Association, in its official statement on artificial intelligence, deliberately draws this same distinction. It says we should think of these technologies not as an “artificial intelligence” that replaces the human, but as “augmented intelligence” — something that strengthens the dentist's clinical judgment, rather than replacing it. In other words, the profession's own governing body says the final decision is always yours, and these tools are meant to help you think better, not to think in your place.

So the principle you take away from this chapter is this: the model is a companion to your thinking, not a substitute for your judgment. And anything that can be fact-checked — a number, a dose, a source — you should assume might be fabricated until you yourself have verified it.

Now, one last point, which is also a door into the next chapter. You might think the solution is simple: just give the model sources so it doesn't have to make things up. And you'd be right; there are tools that do exactly that, like NotebookLM, which speaks only from the sources you give it and builds far less from itself. But this raises a question: if the solution is this simple, why aren't even these tools entirely immune to error? The answer to that — and how to actually work with the model so these errors become rarer — is where the second chapter begins.