The Dividing Line

After a few weeks of working with the model, one question usually stays in your mind: what is this actually good for? Not in general — specifically, in our everyday professional life as dentists.

If you go looking for the answer, you’ll find a pile of lists on different sites. “Ten things AI does for dentists.” “Twenty things every dentist should know about AI.” Memorizing these isn’t bad — you pick up a few good ideas. But next week a task comes up that wasn’t on that list of twenty, and you’re back where you started.

What actually helps isn’t a list, it’s a criterion — something you can hold up against any new task yourself and see whether this one belongs to the model or not.

In its simplest form, the criterion is this: the model works with words, not with the patient.

And the practical way to apply it is a single question. Before any task, ask yourself:

Do I already have the information this task needs, or am I expecting the model to bring it itself?

If you have the information and the model is only going to change its form, that’s the safest case there is. You’ve written a text and want it simplified. You have a paper and want to understand it. You’ve written a messy note and want it tidied up. In all of these, the model adds nothing to what you already have — it only rearranges it. And if it rearranges it badly, the original is right in front of you, and you’ll notice.

But if you don’t have that information and you’re waiting for the model to bring it, you’re standing on the other side of the line. There, the model produces something you have no original to compare it against. And Chapter 2 showed that exactly this kind of fabrication looks confident, fluent, and flawless.

The Bright Side of the Line

Four tasks that sit squarely on this side:

Post-surgical instructions in the patient’s language. You know what to say after flap surgery. The problem isn’t not knowing; it’s that thirteen times a day you have to say the same thing in a way an anxious patient understands and still remembers at home. Write the text yourself, and let the model rewrite it.

Chairside notes. Let me give an example from my own work. Part of DentCast’s clinical content is written exactly this way: chairside, with the patient in the chair, I jot down whatever caught my interest right there — a few broken, scribbled lines that I can barely read myself the next day. Then I sit down and tell the model the same thing, with the details of what I actually saw. The model adds nothing to the account; it edits. The content came from the chairside, not from the model’s memory.

Notice how different this is from “write me a clinical text about this topic.” In the first, you’re organizing something you actually saw; in the second, you’re asking the model to guess what you saw.

Working through a paper. When I want to read a paper for DentCast, the method matters more than the tool itself. I give the model the paper itself, and then we talk about it together. It’s not that I sit across from it like an all-knowing professor and ask, “what do you know about this paper?”

The difference isn’t only in accuracy, it’s in my role. When I’ve read the paper and the model says something, I push back on it. And because I have my own knowledge of the paper, I know right there where it got something wrong. Chapter 2 said the model can misread a real source; this back-and-forth is exactly where the misreading gives itself away. After a few rounds, I end up with a text that’s both fluent and clear — and more importantly, I know it’s correct, because I’ve checked it step by step.

And one that gets thought about less: put the model across from you, not beside you.

Write out a treatment plan you’ve already put together, and tell the model to point out its weaknesses, to say what you might have missed, to say what a demanding colleague would object to.

This is safe for one specific reason: the plan is your own. The model doesn’t make the decision; it only offers an angle that your tired seven-o’clock-in-the-evening mind has lost. And more importantly, you’re able to judge its answer. If it raises an irrelevant objection, you’ll know; if it points out something genuinely missed, you’ll know that too.

The difference from asking “what plan should I give this patient” is subtle, but everything hinges on it. In the first, you’ve made the decision and are asking for critique. In the second, you’ve handed the decision over.

Which brings us to the question that’s been left hanging so far: so what is the model’s role in diagnosis? Now that models can read radiographs too, where does that leave us?