Lesson objective
Explain in plain terms how a language model is trained and generates text one token at a time, and convert each part of that mechanism into a checking habit for dental practice work.
Plain-language explanation
Why this matters now
Most people in a dental practice meet a language model through a chat box, or through a button inside software they already use: "draft this message", "summarise this document", "explain this to the patient". The interface makes the system look like a very well-read colleague who remembers everything. Knowing roughly what happens after you press enter lets you predict where it will go wrong, and that is part of doing the job well rather than a technical hobby.
It also matters because of where dental teams say they want to use these tools next. Among HPI's respondents who are not yet using AI, the tasks they most often plan to adopt are administrative: charting and note taking 34.8%, insurance verification 32.6%, and imaging and diagnostics 25.4%. Those first two are text tasks, and text tasks are exactly where a language model's mechanism decides what can go wrong. As always, those figures are from HPI's Q2 2026 poll of 552 private-practice respondents, invitations sent 15 June 2026, self-reported and not weighted to all US dentists, with HPI's own warning that the panel skews toward ADA members, owners, general practitioners and urban practices.
A question to sit with
If you asked the same assistant the same question about a patient's benefit plan on Monday and again on Tuesday, and it gave you two different answers, which one would you put in the file, and how would you defend that choice?
What goes wrong
Treating the model as a database. People ask a model for a figure the way they would look it up in a booklet. A general language model does not retrieve a stored fact unless the product has been connected to a source. It produces text that is likely to follow your question.
Assuming it knows this year's rules. Training stops at a point in time. A plan's frequency limitation, a state board rule, a form, a code-set year: anything that changed after that point is simply not there, and the model will not necessarily tell you it does not know.
Reading the explanation as evidence. The neat step-by-step reasoning a model shows is produced the same way as the answer itself. It can be just as wrong, and it makes a wrong answer look checked.
Expecting the same answer twice. The model builds each reply from likely continuations, and many products allow variation in which continuation is chosen. Two runs of one prompt can differ in wording, in structure and sometimes in substance.
Forgetting that pasted text is also input. Everything you place in front of the model — your instruction, a referral letter, a benefits booklet, an email from a supplier — becomes part of the same stream of text it is predicting from. Wording inside a document can steer the result.
The practical answer
You do not need the mathematics. You need four plain ideas, and one habit attached to each.
1. Training. A model is built by exposing it to very large amounts of text and adjusting it, over and over, to get better at predicting what comes next. What it keeps is not a library it can open, but a very large set of learned patterns about how words, numbers and ideas tend to follow one another. Developers usually refine it further so that it follows instructions and holds a conversation. Training ends at a date.
2. Tokens. The model does not read letters and words the way you do. Text is broken into tokens: short pieces that may be a whole word, part of a word, a run of digits or a punctuation mark. A figure such as 1,248.50 may be split into several tokens. To the model a number is a sequence of text pieces, not a quantity it is holding in mind.
3. Prediction. When you send a prompt, the model predicts a likely next token, adds it to the text, and predicts again, repeating until the reply is finished. Every answer is assembled one piece at a time from patterns. Nothing in that loop compares the answer with the world.
4. One stream of text. Your instruction, the document you pasted and the model's own reply are all tokens in the same window. There is no wall between "what I was told to do" and "what I was given to read". This is also why security people treat documents a system reads as untrusted input; Lesson 3 returns to it.
Now attach the habits.
| What the model is doing | What it means in a dental practice | The habit |
|---|---|---|
| Predicting likely text from patterns | A likely-sounding answer and a correct answer are different things | Verify content against a source you are holding |
| Working from training that ended at a date | Plan terms, board rules, forms and code-set years may be stale | Supply the current document; never ask it to recall a rule |
| Treating numbers as pieces of text | Arithmetic can go wrong with no sign that it has | Recompute every figure outside the tool |
| Choosing among likely continuations | Two runs of one prompt can differ | Save the output you relied on, with the prompt, tool and date |
| Reading everything in the window as one stream | Wording inside a document can steer the answer | Treat supplied documents as data and check the reply against them |
| Generating its explanation the same way as its answer | The reasoning can be as wrong as the conclusion | Test the conclusion against evidence, not against the explanation |
Supply; do not ask it to recall. The single most useful change follows straight from the mechanism. Compare two requests about the same task.
- Weak: "What is the frequency limitation for adult prophylaxis on the Harborlight plan?" The model has never seen that plan. It will produce a plausible sentence.
- Stronger: "Using only the plan booklet pasted below, dated 1 January 2026, state the frequency limitation for adult prophylaxis and quote the sentence that gives it. If the booklet does not say, reply exactly: not stated."
The stronger request does not make the model reliable. It makes the answer checkable, because you hold the source and you have asked the model to point at it. Module 6 builds this into a full prompt structure. Module 5 covers how to confirm that a rule you found is still in force.
Recompute, do not re-read. Because numbers are text to the model, a treatment estimate, an outstanding balance or a payment schedule must be recomputed in the practice-management system or a spreadsheet. Reading the model's arithmetic back to yourself is not a check; it is the same sentence twice.
Save what you relied on. Because two runs can differ, "the assistant said so" cannot be reconstructed later. Keep the prompt, the output you actually used, the tool and the date with the work.
Use a shared vocabulary when you review. NIST AI RMF 1.0 (26 January 2023) describes trustworthy AI as valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair with harmful bias managed. The framework is voluntary and binds no dental practice, and NIST states it is being revised as part of the White House AI Action Plan, with no revised version published as of 15 September 2026. Three of those words become practical review questions: valid and reliable (does the output match the document?), accountable and transparent (can we show what was asked and who accepted it?), and explainable and interpretable (can the person who used it say why the answer is right, without leaning on the model's own explanation?).
Remember this
- A language model predicts likely text one token at a time; nothing in that loop checks the answer against the world.
- Training ends at a date, so supply the current booklet, rule or form rather than asking the model to remember it.
- Numbers are pieces of text to the model: recompute every figure somewhere else.
- The model's explanation is generated the same way as its answer and can be equally wrong.
- Save the prompt, the output, the tool and the date, because the next run may differ.
Dental practice example
Larkspur Dental Group PLLC is a fictional three-dentist practice. Benefits coordinator Priya Raghunathan is preparing a treatment estimate for an invented patient, Ms Delphine Achebe, whose plan is the fictional Harborlight Preferred booklet sitting in the file. On Monday, Priya types the plan name into the practice's approved assistant and asks for the frequency limitation on adult prophylaxis and the annual maximum. The reply is immediate and confident: two cleanings per calendar year and an annual maximum of $1,500.00, with a tidy sentence explaining that this is "standard for preferred plans". Nothing was looked up. The booklet in the file states a limit of two cleanings in any twelve-month period measured from the last date of service, and an annual maximum of $1,250.00 on a plan year beginning 1 July. Both of the assistant's figures would have gone into an estimate the patient relied on. On Tuesday, Priya tries again the right way round. She pastes the booklet text and asks the assistant to state the frequency limitation and the annual maximum using only the pasted document, quoting the sentence for each, and replying "not stated" if the document is silent. This time the quoted sentences match the booklet, and Priya reads both sentences in the booklet herself before using them. She then runs the same prompt a second time to produce a clean copy for the file and notices the reply is worded differently and rounds one figure; the quoted sentences are still right, so she keeps the first output and records that the second differed. She builds the estimate arithmetic in the practice-management system rather than accepting the assistant's total, and the resulting $612.00 patient portion is recomputed from the fee schedule and the plan year, not from the chat window. In the file she records the prompt, the output she used, the tool, the date, and her own name as the administrative owner of the estimate. The one sentence the assistant drafted that began "because your gum condition" is deleted; anything clinical goes to Dr Rowan Feltrin, who owns it.
Responsible-use warning
Do not ask a language model to remember a plan term, a board rule, a fee or a form and then treat the reply as an answer. It has not looked anything up; it has produced text that is likely to follow your question, from training that ended at some past date. A confident number with a neat explanation is exactly what confabulation looks like. Give the model the document, ask it to quote the sentence that holds the answer, then open the document and read that sentence yourself. Recompute every figure outside the tool, and save the prompt, output, tool and date.
Human-review requirement
The reviewer confirms that no plan term, rule, fee or date in the finished work came from the model's memory: each traces to a document held in the file or to a recomputation outside the tool. The reviewer checks that the prompt, the output relied on, the tool and the date are saved, and that the model's own explanation was not treated as evidence. Ownership follows the work: the practice manager or named administrative owner for an estimate, letter or schedule; the billing staff member, working from the treating dentist's record, for a claim; and the licensed dentist for anything that conveys clinical information to a patient.
Practical exercise
Watch the mechanism work, on invented material only. Write your own fictional document for this exercise. Put no patient information, no real plan booklet and no real record into any tool.
- Write a short fictional benefits booklet extract of about 150 words for an invented plan. Include one frequency limitation, one dollar maximum and one effective date.
- Without pasting it, ask your practice's approved assistant what that invented plan covers for adult prophylaxis. Record the reply word for word.
- Now paste the extract and ask for the same answer using only the pasted text, quoting the sentence that states it and replying "not stated" if it is absent. Record the reply.
- Run the step 3 prompt twice more. Note every difference in wording, structure or substance across the three replies.
- Ask for a calculation that uses your dollar maximum, such as the remaining benefit after two invented claims. Recompute it yourself and note whether the tool was right.
- Ask the tool to explain its calculation. Decide whether that explanation would have helped you catch the error, and write one sentence on why or why not.
- Build a two-column table: for each of the four ideas in this lesson — training, tokens, prediction, one stream of text — write what you observed and the habit it supports.
- Record the prompts, the outputs, the tool and the date exactly as you would if this had been real work.
Knowledge check
-
In plain terms, how does a language model produce an answer?
Answer guidance: It breaks text into tokens and, using patterns learned during training, predicts a likely next token, adds it, and predicts again until the reply is complete. Nothing in that loop compares the answer with a source.
-
Why is asking a model to recall a plan's frequency limitation a poor practice?
Answer guidance: The model does not look it up. It generates likely text from training that ended at a past date, so a current or plan-specific term may be missing or invented. Supply the booklet, ask it to quote the sentence, and read that sentence in the booklet yourself.
-
What does NIST AI 600-1 say about the explanations generative models give, and what does it not say?
Answer guidance: It says outputs may include confabulated logic or citations that purport to justify the answer, and that models sometimes give logical steps even when the answer is incorrect. It does not quantify how often confabulation occurs, and it is voluntary and cross-sectoral rather than dental.
-
Two runs of the same prompt give different replies. What should be in the file?
Answer guidance: The prompt, the output actually relied on, the tool and the date, together with the source document the answer was checked against. Because the next run may differ, "the assistant said so" cannot be reconstructed later.
Sources and further reading
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.
- American Dental Association, Health Policy Institute — Dentists Use AI to Make Appointments More Efficient, but Draw the Line at Clinical Decision Making (HPI Perspectives) (2026-07-01). Accessed 2026-09-15.
- American Dental Association, House of Delegates — Policy Statement on the Use of Augmented Intelligence in Dentistry (Trans.2021:282), as printed in ADA Current Policies, Adopted 1954-2025. Accessed 2026-09-15.