Would You Let Software Tell a Patient What They Need?
The American Dental Association's Health Policy Institute put this to a national panel of 589 dentists in the second quarter of 2026, and the answer has a very particular shape. Just over four in ten use AI for at least one task. Fewer than one in twenty use it for patient treatment recommendations, and 82.6% say they do not plan to. The same shape appears in medicine, where a survey of 1,692 physicians found 81% using AI in practice while assistive diagnosis sat at 17%. That gap is not timidity, and it is not a technology-adoption problem waiting to be solved. It is a profession drawing a line in exactly the right place, and the AI Alignment Filter™ ends on that line: is this tool replacing judgment, or enhancing it?
Introduction
There is a story the technology trade press likes to tell about dentistry, and it is that we are slow.
I have been hearing it for fifty years. It was said about digital radiography, about intraoral scanners, about electronic records. Sometimes it was fair. Often the people saying it were selling the thing we were supposedly slow to buy.
So when survey data arrives showing that dentists use AI enthusiastically for some things and refuse it flatly for others, there are two available readings.
One is that adoption is incomplete, and the remainder is a marketing problem.
The other is that a profession of licensed clinicians has looked at a class of tool and sorted it correctly, on the first pass, without being asked to.
I think the second reading is right, and this week is the case for it.
Key Takeaways
• Among 589 dentists surveyed by the ADA Health Policy Institute in Q2 2026, 43.3% used AI for at least one task and fewer than 5% used it for patient treatment recommendations.
• 82.6% say they do not plan to use AI for treatment recommendations. That is different from saying they never will, and the distinction matters.
• Medicine shows the same shape. Among 1,692 physicians surveyed by the AMA in 2026, 81% used AI while assistive diagnosis sat at 17%.
• The adoption is concentrated where the work is administrative, and the reluctance is concentrated where the work is clinical judgment. That is not an accident.
• Physician adoption ran 38% in 2023, 66% in 2024, and 81% in 2026. The shape is a steep jump followed by a marked slowdown, not a straight line upward.
• The line is not caution. It is the correct answer, and the frameworks exist to keep it drawn deliberately rather than by instinct.
The Shape of the Number
Take the ADA Health Policy Institute's Q2 2026 panel, 589 dentists, reported in July of that year.
43.3% use AI for at least one task in the practice. That is not a profession refusing a technology.
Fewer than 5% use it for patient treatment recommendations.
And 82.6% say they do not plan to.
Where the use is concentrated tells you the same thing from the other direction. Charting and note-taking is the largest planned-adoption category among those the ADA reported, at 34.8%. Insurance verification is in use by 13.6%, with 32.6% planning to adopt it.
So the picture is a profession that will hand a machine the periodontal charting, the notes, and the insurance verification, and will not hand it the question of what this patient needs.
It is also a profession reluctant to put AI between itself and the patient in conversation. 68.3% say they do not plan to use AI for explaining clinical findings to patients.
That is a coherent position, arrived at independently by hundreds of clinicians who were not coordinating with each other, and it maps almost exactly onto the boundary this framework draws.
The Same Shape in Medicine
If dentistry were simply behind, you would expect medicine to show a different pattern. It does not. It shows the same one, at a larger scale.
The American Medical Association's 2026 survey of 1,692 physicians found 81% using AI in practice. Look at what for. Research summarization at 39%. Documentation at 30%. Billing and coding at 28%.
Assistive diagnosis: 17%.
Same shape. Enthusiastic for the administrative, restrained for the clinical.
There is a detail in the adoption trajectory worth pulling out, because it is usually reported in a way that gets the story backward. Physician AI use ran at 38% in 2023, 66% in 2024, and 81% in 2026.
Quoted as 38% to 81%, that reads as one explosive leap and is used to argue that adoption is accelerating. The actual shape is a very steep rise through 2023 and 2024, followed by fifteen points across the next two years. That is deceleration.
It is a small example of this month's whole lesson. The two endpoints were both true, and the story they told with the middle removed was the opposite of the story the data tells.
That Is Not Timidity
Here is why I read this as competence rather than caution.
The tasks clinicians have handed over share a property. They are reversible, they are checkable, and an error in them is visible to a person who knows the work. A misheard pocket depth gets caught at the next appointment. A miscoded claim comes back. A badly summarized article gets read by someone who knows the literature.
The tasks clinicians have kept share the opposite property. A treatment recommendation is consequential; it is often not reversible, and the patient has no independent way to evaluate it. The only safeguard is the clinician's judgment, and the clinician is the one who carries it, legally and otherwise.
A profession that sorts tools by where the error becomes irreversible and unverifiable has not been slow. It has been running the right test.
The concern clinicians raise most about this technology is not that it will fail. It is deskilling: 88% of the physicians in that AMA survey reported at least some concern about skill loss. That is a concern about what happens to judgment when it is not exercised, from the people whose judgment it is.
What "Do Not Plan To" Actually Means
One careful point, because this figure travels badly and I would rather you hear the accurate version from me than the compressed version from somewhere else.
The ADA's published wording is that 82.6% "do not plan to" use AI for treatment recommendations. It is frequently repeated as "82.6% say they never will."
Those are not the same statement. Not planning to do something is a statement about current intention. Never is a statement about the future. The ADA separately reports a smaller never-intend figure, closer to one in three.
Why does this matter enough to spend a paragraph on? Because the compressed version makes the profession sound immovable, and the accurate version describes something much more interesting: a large group of clinicians who are not doing this now, are not committed against it forever, and are presumably waiting for something. Evidence, most likely. Or a product that respects the line rather than trying to erase it.
That is a profession that can be persuaded by evidence and has not yet been. Which is exactly what you would want it to be.
The habit is the transferable part. When a survey figure is quoted at you, ask what the question actually said. The gap between the question and the headline is where most misreading lives.
The Line, Drawn Precisely
The Filter's fourth step asks whether a tool is replacing judgment or enhancing it. In practice, that line can be drawn concretely, and it should be written down rather than carried around in someone's head.
AI may assist with reminders, reactivation, and confirmations. Draft communication. Documentation and SOPs. Review templates. Meeting summaries. Research synthesis. Scheduling and the administrative drag that eats the day.
AI does not touch clinical decision-making or treatment planning. It does not handle financial policy exceptions. It does not respond to sensitive complaints or ethical grey areas. It does not conduct performance conversations, mediate conflict, or do the work of building a culture.
Notice that the second list is not only clinical. The common property is that the person on the other side needs to be dealt with by a person with standing to decide, and a fluent automated response is not a cheaper version of that.
Four lines in a written policy will hold this: where AI is encouraged, where it is limited, what must always be reviewed, and who retains final responsibility. Most practices have none, which means adoption happens quietly and unevenly, one person drafting and another summarizing and nobody discussing it.
The Patient Is the Reason
This is the part that gets lost when the conversation is about efficiency.
If a tool saves your team thirty minutes of drafting, the gain is not the thirty minutes. It is what those thirty minutes become. Sitting longer with the patient who has questions. Not rushing the conversation about a treatment plan that somebody is frightened of.
Automation should protect human interaction, not replace it. That is the exchange, and it only works if you decide in advance what the reclaimed time is for. If you do not, it is absorbed within a fortnight, and the practice is simply faster, which nobody went into dentistry for.
There is a version of AI adoption in this profession that ends with clinicians spending more time with patients and less time with keyboards. It is genuinely available. It requires knowing where the line is and defending it, which is not the same as being cautious.
Frequently Asked Questions
Is any AI product cleared to make treatment recommendations?
Cleared dental AI products are diagnostic aids on radiographic images, and their intended use statements are narrow. A product that marks a suspicious area for a clinician to examine is doing something quite different from recommending treatment, whatever the interface suggests. Read the intended use.
My associate uses AI to draft treatment plan explanations. Is that over the line?
Drafting communication is on the permitted side. What matters is whether a clinician genuinely evaluates it before it reaches a patient, and whether the clinical content originated with the clinician rather than the tool. If the machine decided what the patient needs and the dentist edited the prose, the order has been reversed.
Should I tell patients when AI is used in their care?
Where AI contributes to a clinical assessment, patients are increasingly expecting to be told, and the conversation is easier than people fear. Most patients are reassured to hear that software flagged something and their dentist checked it, because that describes a second look rather than a substitution.
We are a small practice. Does any of this apply to us?
More, not less. A large group has compliance staff who will notice an unwritten policy. In a small practice, the policy is whatever the owner does, and it gets copied.
What would change my mind about treatment-recommendation AI?
Independent, multi-site evidence of patient outcomes, not accuracy against a reader panel. Published limitations and failure modes. And a product that surfaces its own uncertainty rather than answering everything in the same confident register. None of that is unreasonable, and none of it exists yet in this category.
Final Thoughts
Dentists are not afraid of artificial intelligence. The data says so, and so does every conversation I have had about it.
What the data actually shows is a profession that picked up the tools that take work off the day and declined the tools that would take judgment off the clinician, and did it quickly, without a framework, on instinct.
The instinct is right. The reason to write it down anyway is that instinct does not survive a persuasive Tuesday afternoon, and it does not transmit to the associate you hire next year.
So put the four lines in writing. Name where AI helps, name where it does not go, name what always gets read by a person, and name who is responsible when it is wrong.
Cleared is not the same as proven. And a machine that predicts is not a clinician who decides.