When the Text Leaves Your Practice, Who Answers for It?

This is the final part of the book. The previous part was about the things you give the model. Now we turn to the return direction: the text you get from the model and that leaves your practice. When that text reaches the patient, who is responsible for it?

A Sheet That Carries Your Name

After surgery and tooth extraction, you hand the patient the post-operative instruction sheet. You got the text for that sheet from the model last week. It is tidy, it reads well, and it is printed on the practice’s letterhead. One of its sentences reads: “From the very first day, rinse your mouth thoroughly with salt water several times.” The patient does exactly that, and three days later comes back with severe pain and a dry socket.

You yourself know perfectly well that vigorous rinsing in the first twenty-four hours dislodges the clot. Had you read that sentence carefully, you would have paused right there. But the text was fluent, the rest of it was correct, and from the outside that one sentence looked no different from the correct ones. This is the same thing we said in Chapter 2: fluency and correctness both come out of a single process, so you cannot separate them by looking at the text itself.

The patient does not know a model wrote this text, and even if they did it would make no difference to them. You handed them the sheet, with your name on it. The company that built the model is not a party to this either. Every one of these services states explicitly in its terms of use that the output may be wrong and that evaluating it is the user’s job.

The Same Thing You Do With the Laboratory

This situation is not new to us. Every day we take delivery of work we did not make ourselves. The laboratory makes the crown, but before cementation you check the margin, the contacts and the occlusion. If a crown is cemented with an open margin, nobody accepts the explanation that the laboratory made it that way. The laboratory made it, but you accepted it, and responsibility transferred at the point where you accepted it.

The model’s output carries the same verdict, with one difference that makes the job harder. An open margin is shown to you by an explorer and a radiograph, but a wrong sentence inside a fluent text gives no signal at all and is found only when somebody who knows the subject reads the text carefully. So that stage of checking before cementation exists here too and is not removed; only its instrument is different. Its instrument is your own knowledge.

Why You Sometimes Imagine This Check Has Already Been Done

The model has several behaviors that, in people, are signs of exactly this kind of checking. When you see those behaviors, you unconsciously ease off your own scrutiny.

The most common of them is the confidence number. You write “only answer if you are more than 80 percent sure” and at the end of its answer the model writes “Confidence: 90 percent.” There is no measurement behind that number. The model produced it the same way it produced every other sentence: in a context like this one, after an answer like this one, a number like this one was probable. Ask the same question in a fresh conversation and you may get a different number, without anything in the model’s knowledge having changed. And the trouble is not merely that the number is useless. This number puts your mind at ease at precisely the place where you should be suspicious.

The error is in the number, not in the caution. You saw the healthy version of the same request in Chapter 6: “If you do not have the information you need, do not guess — ask me.” That sentence asks the model for a behavior, not for a measurement of its internal state.

At this point the question may arise of why, in some places, getting a number out of a system is the right thing to do. The difference is that those numbers are attached to something outside the model.

One example is when you set the criteria yourself. Suppose that, to judge the quality of a paper, you write out a list of conditions — that the study has a control group, that its sample size is not below a certain threshold, that its follow-up period is specified — and then ask the model to assess a paper against that same list, say for each item whether it is present or not, and score it. The number that comes out at the end is the sum of the scores for those items, and you can look at them one by one and see where it came from. Here the model is judging the paper, not judging itself.

DentCast did exactly this: we wrote a prompt that reviews papers item by item, weighs the evidence and scores them against a template. It is called the DentCast Evidence Score.

Note that here the model is not assessing itself; it is judging a paper against a criterion.

Another example is the confidence percentage that radiographic diagnostic software puts beside each finding. That number was obtained by testing the system on thousands of labeled images, and the manufacturer has to be able to show how it was calculated. But a language model’s “I am 90 percent sure” is attached to nothing at all.

You saw three further behaviors in detail in Chapter 6, and here I will only give their conclusion. An agent that finds and fixes its own error in front of you only tests what gets executed, and a sentence written about a dose or about post-operative care does not pass through such a test at all. Two models that give the same answer share a large part of their training data, so their agreement shows that the statement is common, not that it is correct. And a model that asks precise questions before starting work makes the text more relevant, not more correct.

In all four cases, the work that appears to have been done has in fact not been done. The text remains exactly as unchecked as it would be without these behaviors.

What Has the American Dental Association Said About This?

The American Dental Association (ADA) has a document called ADA SCDI White Paper No. 1106, published in 2022, on applications of artificial intelligence in dentistry. Part of that document is directly about responsibility.

Its position is clear. Artificial intelligence in treatment is only an adjunct to the clinician’s work, and responsibility for diagnosis, prevention and treatment remains with the dentist. To explain this it uses the self-driving car analogy: as long as the driver is responsible for driving safely, they must stay awake with their hands on the wheel, however well the system works. It then warns explicitly that, as long as diagnosis and treatment planning rest with the clinician, one must guard against over-reliance on these systems.

There is one more point I should add, which happens to make this argument stronger. This document was written before ChatGPT arrived, and its subject is not language models but radiographic diagnostic software — the first category we talked about at the beginning of the book. That software is built for one defined task, has been validated on labeled data, and has been cleared by a regulatory body. Even so, the ADA does not lift responsibility off the dentist’s shoulders. Now, a general-purpose language model has none of those three. It was not built for dentistry, it has not been validated for any clinical use, and no body has approved it. When responsibility does not transfer in the case of that validated tool, it transfers even less in the case of this one.

What Does Responsibility Mean in Practice?

Saying “the responsibility is yours” is easy and on its own changes nothing. What does change things is three specific habits.

First, read every text that is going to reach a patient in full, the way you read something written by a new assistant. I do not mean a quick skim to see whether it came out well. I mean reading it sentence by sentence with this question in mind: if the patient does exactly what is written here, what happens? These texts are usually short, and this takes no more than a few extra minutes.

Second, for every clinical recommendation inside the text you must be able to say where it came from, without mentioning the model. If a colleague asks why you prescribed a mouthwash for a week, your answer has to be that this is your own protocol, or that a particular source says so. If the only answer you have is that the model wrote it that way, that sentence is not yours yet and should not go out under your name. This is the same constraint we stated in Chapter 1: the model is a tool only within the territory you yourself know.

Third, keep the text you have checked once and use that same one next time, instead of getting a fresh text from the model each time. The post-operative sheet you have read and corrected is now your own sheet. The text the model writes tomorrow on the same subject is a different text, with different possible errors, and the check you did yesterday does not cover it.

For the scientific content you publish, I described the working method in Chapter 6 and will not repeat it. Only one thing is added to it. Your name under that text means you have said this statement is correct, and the reader is trusting your credibility, not the model’s.

The End of the Book

From the start I stated one principle that was meant to hold throughout the book: the model is a companion to thought, not a substitute for judgment; and every verifiable claim, from a number or a dose to a reference, may be fabricated until it has been confirmed. Each of the other chapters was one facet of that same principle. Why the model speaks with confidence when it does not know. How to talk to it so you get a useful answer. Where in clinical work it has a place and where it does not. How to build content with it that you can stand behind. And in this chapter, what you should not give it and who answers for what you get from it.

The names of the models, their capabilities and their subscription prices will change within a few months, and some of this book’s examples will go stale. What does not change is your own position in this relationship. This tool can write faster than you, read more than you and summarize more neatly than you. What it cannot do is examine the patient, see the outcome of the treatment, and answer for it. Those three are the dentist’s work, and all three stay with you.