Lesson objective
Distinguish rules-based automation, machine learning and generative AI in dental practice work, and match the check you apply to the way each one fails.
Plain-language explanation
Why this matters now
Walk through a dental practice on a normal Tuesday and you will pass four or five things called "AI" before lunch. The practice-management system sends a recall text on its own. A scheduling screen shows a "no-show risk" beside a name. An imaging program puts a coloured box on a bitewing. A chat assistant drafts the letter explaining a financial arrangement. The label on all of them is the same word. The systems underneath are not the same, and they fail in different ways.
How common is this? One dental figure is recorded, and it has to be quoted with its sample. The ADA Health Policy Institute reports that, as of mid-2026, 43.3% of responding dentists use AI for at least one type of task in their practice, and a further 26.4% plan to. Those numbers come from HPI's Economic Outlook and Emerging Issues in Dentistry Poll, second quarter 2026: invitations went to 2,432 panel dentists on 15 June 2026, 589 responded (an adjusted response rate of 24.3% after undeliverable addresses), and the AI results rest on the 552 respondents who work in private practice. HPI warns in the deck itself that this quarter's respondents skew toward ADA members, general practitioners, practice owners, solo or small-group practice, mid-to-late career, white dentists and urban areas. It is an opt-in panel, not a probability sample of US dentists, and HPI notes that other published research finds larger take-up. There is no verified earlier dental baseline in this programme's register, so nobody on this course may describe the figure as a rise or a trend.
What the figure does establish is that a dental team member today is far more likely to be using one of these systems than to have been taught what kind of system it is. That is the gap this lesson closes.
A question to sit with
Your software sends a recall text on its own, scores a patient for no-show risk, and drafts a letter about a payment plan. All three are described as AI. Which of them could be wrong in a way you would never notice just by reading what it produced?
What goes wrong
Treating one word as one thing. A practice writes a single line in its policy, "staff must check AI output", and applies it identically to a recall rule and to a generated letter. The rule gets re-read every month and the letter gets skimmed every time.
Checking generated text the way you check a rule. A rule that was tested once behaves the same way the next time it meets the same facts. A generative system writes new text on every run. Having checked yesterday's letter tells you very little about today's.
Reading a score as a verdict. Machine-learning features often show a number beside their suggestion. That number is the system's own estimate of how strongly the patterns it learned point to that answer. It is not evidence about this patient, and it is weakest exactly where the case is unusual.
Mistaking fluency for understanding. A well-written paragraph about a patient's benefit plan can rest on a misread limit or an invented clause. Fluency is a property of the text, not of the reasoning behind it.
Letting the product category decide who owns the result. No tool moves responsibility. The person who would have owned the work without the tool still owns it with the tool.
The practical answer
Sort every AI feature you meet into one of three families before you decide how to check it. Products mix families, so sort the feature, not the box it came in.
| Rules-based automation | Machine learning | Generative AI | |
|---|---|---|---|
| How it works | A person writes explicit if-then logic | The system learns patterns from past examples | A model trained on very large amounts of text or images produces new output |
| Where a practice meets it | A recall rule that texts every patient six months after a hygiene visit; a form that will not submit without a date of birth | A no-show risk score; a flag on a radiograph surface; a suggested benefit category | Drafting a recall message, a financial-arrangement letter, a post-operative instruction sheet for the dentist to approve, a summary of a long benefits document |
| How it fails | Consistently: the rule is wrong, or the case falls outside it | Quietly, on the unusual case; and it drifts as the practice changes | Plausibly: confident, well-formed content that is false |
| What a wrong output looks like | The same wrong result every time | A reasonable-looking suggestion, often with a score | A fluent, complete-looking answer |
| How you check it | Test the rule against real cases, review its exceptions, control who may change it | Sample the suggestions, look hardest at unusual and low-score items, watch for drift | Verify every fact, figure, date and reference against a source you hold |
Three questions turn that table into a habit.
1. Which family produced this? If you cannot tell from the screen, ask the vendor or whoever administers the system, and write the answer down. "Unknown" is an acceptable interim answer; guessing is not.
2. How does that family fail? Match the check to the failure. A rule needs a test. A suggestion needs sampling. Generated text needs line-by-line verification.
3. Who owns the output? This depends on the work, not on the tool.
| The work | Who owns it |
|---|---|
| Diagnosis, the reading of images, the treatment plan, the prescription, the clinical record, any patient communication that carries clinical information | The licensed dentist |
| A delegated task, within the duties that person's state permits and at the supervision level that state requires | The hygienist or assistant performing it |
| The practice's compliance with HIPAA and board rules | The practice owner or dentist of record, with the designated privacy official and security official |
| A claim, prepared from the treating dentist's own record | The billing or dental-benefits staff member; the code must be supported by what the dentist documented |
| Schedules, recall and reactivation messages, forms, non-clinical letters, operational reports, vendor records | The practice manager or the named administrative owner |
Sorting by family also tells you where the check has to sit in your day. A rule is checked on a schedule, because nothing prompts you. A suggestion is checked when it is acted on. Generated text is checked before it leaves the building. Write the three checks into the workflow rather than into a person's memory.
One more habit belongs here because it is the hardest to retrofit. Human roles have to be designed, not assumed. The NIST Artificial Intelligence Risk Management Framework (AI RMF 1.0, NIST AI 100-1, released 26 January 2023) says in Appendix C that human roles in decision making and oversight "need to be clearly defined and differentiated", and that human-AI configurations can span "from fully autonomous to fully manual". Appendix C also notes that under some conditions the AI part of a human-AI interaction "can amplify human biases, leading to more biased decisions than the AI or human alone". The framework is voluntary. It is non-sector-specific, it binds no dental practice, no US regulator requires a practice to adopt it, and there is no NIST certification to hold. As of 15 September 2026 NIST states that AI RMF 1.0 is being revised as part of the White House AI Action Plan, and no revised version had been published.
Finally, be honest about the limits of what the register records. No dental AI taxonomy or category scheme has been published that this course can point to. The three families above are a teaching device for deciding how to check something, not an official classification, and the next lessons keep them.
Remember this
- "AI" covers three families — rules-based automation, machine learning and generative AI — and they fail in different ways.
- Match the check to the failure: test the rule, sample the suggestions, verify generated text line by line.
- A score beside a suggestion is the system's own estimate from past patterns, not evidence about this patient.
- Quote the HPI adoption figure only with its sample (552 private-practice respondents), its date (June 2026) and HPI's own warning about skew; there is no verified baseline, so claim no trend.
- The tool never moves ownership: the licensed dentist owns anything clinical, and a named administrative owner owns the rest.
Dental practice example
Cedarbrook Family Dentistry is a fictional four-operatory practice with two dentists and eleven staff. Office manager Nadia Brennan is asked to list every feature in the practice that is sold as AI and say how each is checked. She finds four. The first is the recall automation in the practice-management system: it texts any patient 5 months and 3 weeks after a hygiene visit. Nadia tests it against last month and finds it texted eleven patients who had already been seen, because the rule reads the last scheduled visit rather than the last completed one, and it texted two patients whose records are marked transferred out. That is rules-based automation, failing consistently, so she has the rule corrected and adds a monthly exceptions report. The second is a no-show risk score on the scheduling screen. It scored a long-standing patient, Mr Idris Kovac, as high risk because his three most recent appointments were rescheduled; all three were rescheduled by the practice when a dentist was out. Nadia stops the front desk from double-booking on the score alone and asks that any score above the threshold be read together with the note on the account. That is machine learning, failing quietly on an unusual case, and the score reflected a pattern rather than this patient. The third is a generative assistant that drafts financial-arrangement letters. Its draft for the Vasquez family stated that "your plan covers two cleanings and one fluoride treatment per calendar year". Nobody supplied the plan document; the sentence was generated because it is the kind of sentence that follows. The actual plan booklet in the file says something different. Nadia rewrites the letter from the booklet and makes supplying the document part of the task. The fourth is an imaging assist that draws boxes on bitewings. Nadia writes one line for it and stops: every box goes to Dr Imani Osei, who reads the images and owns the finding, and no box is shown to a patient or written into a treatment presentation before that. Her one-page table names each feature, its family, its check, and its owner: Nadia for the recall rule, the schedule and the letters, and Dr Osei for anything the imaging assist produces.
Responsible-use warning
Do not let a high score or a polished paragraph shorten your review. A score is the system's own estimate from patterns it has seen before, and it is least reliable on the unusual patient who matters most. A fluent generated letter can state a benefit limit, a frequency or a date that no document in the file supports. Neither is evidence. Before you accept an AI output, name the kind of system that produced it and apply the check that matches its failure: test the rule, sample the suggestion, or verify every fact and figure in the draft against a source you are holding.
Human-review requirement
Who reviews depends on the work. Anything clinical — a finding on an image, a diagnosis, a treatment plan, or a patient communication that carries clinical information — is read, corrected and owned by the licensed dentist, and ADA policy states that an AI system in clinical dental practice should be supervised by a dentist. Administrative output is owned by the practice manager or the named administrative owner, and a claim is owned by the billing staff member working from the treating dentist's own record. The reviewer confirms the family of each feature was identified, that the check matched its failure mode, and that nothing was accepted on the strength of a score or fluent wording alone.
Practical exercise
Map the AI features in one workflow you know. Use fictional or invented examples only, and put no patient information into any tool for this exercise.
- Choose one recurring task in your practice: a week of recall messages, a morning of schedule filling, a batch of benefit summaries, or the new-patient intake.
- List every feature in the software you use for that task that is described as AI, smart, automatic, predictive or suggested.
- For each feature, decide its family: rules-based automation, machine learning or generative AI. If you cannot tell from the screen, write "unknown" and the name of the person you will ask.
- For each feature, write one sentence on how that family fails.
- For each feature, write the check that matches that failure: a rule test against real cases, a sample of suggestions weighted toward unusual ones, or line-by-line verification of generated text.
- Build a four-column table: feature, what the system does, what a person does, who decides.
- Name the owner of each output using the categories in this lesson — licensed dentist, hygienist or assistant within their state's permitted duties, practice owner or dentist of record, billing staff working from the dentist's record, or practice manager.
- Mark any feature where the current check does not match the family, and write one change you would take to the practice owner.
- Keep the table with your workbook notes. You will reuse it when you build the practice AI risk matrix later in this Level.
Knowledge check
-
Why can a recall rule and a generated patient letter not be checked the same way?
Answer guidance: The recall rule is explicit logic that behaves the same way every time, so it is tested against real cases and its exceptions reviewed on a schedule. A generative system writes new text on every run and fails plausibly, so every fact, figure and date in each output must be verified against a source before it is sent.
-
What does a no-show risk score tell you, and what does it not tell you?
Answer guidance: It is the system's own estimate of how strongly the patterns it learned point to that answer. It is not evidence about this patient, and it is least reliable on the unusual case, so it is read alongside the account notes rather than acted on alone.
-
What must you state whenever you quote the HPI figure that 43.3% of responding dentists use AI for at least one task?
Answer guidance: That it comes from HPI's Q2 2026 poll of 552 private-practice respondents, with invitations sent on 15 June 2026 and an adjusted response rate of 24.3%; that HPI warns the respondents skew toward ADA members, general practitioners, owners, solo or small-group, mid-to-late career and urban dentists; and that no earlier verified baseline exists, so no trend may be claimed.
-
What is the status of the ADA policy statement on augmented intelligence, and what does it say about supervision?
Answer guidance: It is ADA House of Delegates policy (Trans.2021:282) stating the association's position. It is not a law or regulation and not part of the Code of Professional Conduct, so it is not enforceable through member discipline as written. It states that AI should supplement rather than replace the dentist's clinical judgment and that an AI system in clinical dental practice should be supervised by a dentist.
Sources and further reading
- 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, Health Policy Institute — The State of the U.S. Dental Economy, 2nd Quarter 2026 Update — Plus: AI Usage in Dental Practices (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.
- 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.