Almost nobody checks, and the reason is not negligence. A survey of 163 dental practice owners in one German state found that the barriers to adopting AI were additional cost, reported by 91 of them, and fear of additional workload, reported by 68. Dentists are not frightened of this technology. They are frightened of paying for something that makes Tuesday harder, and a governance routine that adds work confirms exactly that fear. So the standard has to make the tool cheaper to live with, not merely safer. That means a short, scheduled check rather than an audit: has anything changed underneath it, is anyone working around it, and would we know if it were wrong? The AI Alignment Filter™ decides whether a tool comes in. This decides whether it has quietly stopped earning its place.
Introduction
This is the last article of the month, and I want to close it where the evidence actually points rather than where the conversation usually goes.
The story told about dentists and technology is that we are cautious. That the barrier is trust, that we need reassurance, that adoption is a matter of overcoming clinical conservatism.
When somebody goes and asks, that is not what comes back.
A study published in Scientific Reports in August 2026 surveyed 163 dental practice owners in one German state about what stops them from adopting AI. The leading barrier was additional cost, named by 91 of them. Second was fear of additional workload, named by 68.
Not distrust. Cost and workload.
That changes what a governance routine has to look like. If the binding constraint is that technology already makes the day harder, then a review process that adds an hour of administration to your quarter is not a solution. It is a confirmation of the problem.
So everything that follows is built to be short.
Key Takeaways
• Among 163 German dental practice owners surveyed in 2026, the leading adoption barriers were additional cost, named by 91, and fear of additional workload, named by 68.
• The binding constraint is not distrust, which means governance has to be cheap in time, or it will not happen.
• Tools drift through software updates, equipment changes, staff turnover, and patient-mix changes. None of those announce themselves.
• A model's published accuracy was produced on its developer's data. Your patients are the external cohort, and performance there is always lower.
• A quarterly check of four things and an annual check of four more is proportionate for almost any practice.
• No new framework is being introduced to do this. The instrument you already have is enough.
A Number That Was Not There
Before the substance, a short account of something that happened while this month was being prepared, because it is the month's own lesson landing on the month's own research.
The source material behind these four weeks included a summary of that German study. It reported three barriers: cost at 57%, workload at 43%, and lack of trust at 13%.
The first two figures were close to right and slightly off. The actual paper reports 91 of 163 respondents for cost, which is 55.8%, and 68 for workload, which is 41.7%.
The third figure does not exist. Not the number, and not the category. The paper's third barrier is limited familiarity with AI applications, and it is reported without a percentage at all.
Now, the conclusion that the summary drew, that cost and workload rather than distrust are the binding constraints, is probably correct. It is consistent with the ADA data and with every conversation I have had on the subject. But the evidence offered for it was a figure nobody had measured.
I could have published that. It was in a well-researched brief, attached to a real citation, pointing at a conclusion I already believed. That last part is what makes it dangerous.
This is why the checking has to be a routine rather than an instinct. Instinct does not catch the errors that agree with you.
What Drift Looks Like in an Operatory
Software in a practice does not degrade the way equipment does. There is no rattle, nothing wears out, and nothing gives you a warning. What changes is the world around it.
The imaging hardware changes. A new sensor, a different panoramic unit, even a change in exposure settings alters the images a detection model sees. That model learned on a particular distribution of images. Yours just moved.
The software updates. Vendors adjust thresholds, change defaults, and enable new features. These arrive in release notes nobody in a dental practice has time to read. Sensitivity can shift without anyone in your building making a decision.
The patient mix changes. You take on a new employer plan, a nursing home, a pediatric cohort. The population the tool is now seeing is not the population it was validated against, and it is not the population you were seeing last year either.
The people change. A new hygienist calibrates differently. A new associate accepts more of the flagged findings, or fewer. The tool has not changed, and the practice's use of it has.
None of these produce an obvious failure. They produce a slow divergence between what the tool does and what everybody believes it does, and the only thing that surfaces it is somebody looking on purpose.
Your Patients Are the External Cohort
There is one research finding from this month that I would keep above all the others, because it converts directly into a purchasing and governance principle.
A model predicting periodontal treatment response performed strongly on the data it was developed on and noticeably worse when tested on a different population. That is not a scandal or a badly built model. It is the ordinary behavior of every predictive model ever made. A model learns the population it was trained on, including that population's particular mix of disease severity, imaging equipment, and recording habits, and some of that does not transfer.
Here is the translation. Any accuracy figure a vendor gives you was produced somewhere else, on somebody else's patients.
Your practice is the external validation. Your performance will be lower than the published figure, and the gap is not a defect. It is what external validation means.
Which is exactly why the check matters. Nobody else is going to measure how this tool performs in your building on your patients. If you do not look, that number does not exist anywhere.
The Check, and How Short It Is
This is a routine, not an audit, and it is deliberately small. No new framework, no scoring instrument, no worksheet. The month introduced no new mark for this on purpose, because the discipline you already have is sufficient, and another named tool would be one more thing to carry.
Four questions, quarterly. Fifteen minutes in a staff meeting.
Has anything changed underneath it? New sensor, software update, new plan, new team member.
Is anyone working around it? Ask it as if you're looking for where the system failed the team, not the reverse.
Has anything reached a patient without a person reading it? This is the one with consequences, and it should be asked in that register.
Is the thing it replaced genuinely still gone?
Four more, annually. An hour, once.
What has it actually cost, including the time nobody counted?
Would we buy it again today, at today's price, knowing what we now know?
If it started producing subtly wrong output next month, what would surface it?
Who owns it, and do they know they own it?
That is the entire program. Fifteen minutes four times a year and an hour once. If governance costs more than that in a practice of any normal size, it will be abandoned within a year, and an abandoned routine is worse than none because it creates a record that stops.
The Question With Consequences
Of the eight above, one is not an efficiency question and should not be asked in that tone.
Has AI-generated output reached a patient without a person reading it?
Everything else on the list is about whether you are getting value. That one is about whether something went out of your building carrying your name and your clinical authority that nobody in your building read.
It happens more easily than people expect. A drafted recall message, a summarized finding, a templated explanation. Each of those is entirely legitimate with a person in the loop and entirely different without one, and the transition between the two states is usually nobody's decision. It is a busy week and a convenient default.
Ask it plainly, ask it in a register that makes the answer safe, and ask it every quarter.
Frequently Asked Questions
How do I check a diagnostic tool's performance in my own practice?
You are not running a validation study. Keep a simple tally over a month of flagged findings you agreed with and flagged findings you did not, and compare it against the same exercise a year later. Direction of travel is what you are after, not a precise figure.
Our vendor says the model improves over time. Does that change the review?
It makes it more necessary, not less. A model that updates is a model that can change behavior between one quarter and the next, and "it improves" is a claim about the average rather than about your practice.
What if the quarterly check keeps finding nothing?
Then it is working, and it costs you fifteen minutes. Stability is the expected result. The check exists for the quarter where something has changed, and you cannot know in advance which one that is.
Should this go in the practice manual?
Yes, alongside the four lines about where AI may and may not be used. Both are short, and the point of writing them down is that they survive the person who set them up.
We adopted several tools this year. Do they all need this?
The ones that touch clinical information or reach a patient, yes. The scheduling tool and the accounting software can live on the annual list. Proportion the effort to what an undetected error would cost.
Final Thoughts
Four weeks, and the argument compresses into four sentences.
Cleared is not the same as proven, and the record is public if you look.
Prediction got cheap, which made judgment expensive, so know which part of your work is which.
Ask who measured it, how many, where, and what else they found.
And a tool nobody owns is a tool nobody checks.
There is no new framework in any of that. The AI Alignment Filter™ has been in use since March and it did not need replacing, which is the honest reason this month minted nothing new. If it does not replace or compress something, it may be unnecessary, and that applies to frameworks as much as to software.
What I would leave you with is the thing that took me fifty years to be sure of. Every technology I have watched arrive in this profession was adopted best by the people who understood their own work most clearly. Not the earliest adopters and not the most cautious. The ones who could say precisely what the difficulty was before anyone showed them a solution.
That has not changed, and this technology has not changed it.
Name the constraint. Ask who measured it. Keep the judgment. Check it twice a year. See you next month.