It is a fair question, and it has a public answer, which is why it is worth asking out loud. Funding does not make research wrong, but it shapes what gets measured, what gets reported, and which of the three numbers reaches the slide. A randomized trial of AI-supported orthodontic triage reported sensitivity of 1.0 and overall accuracy of 0.98. It also reported specificity of 0.67, meaning one in three patients who did not need a referral was flagged as needing one. Two of those figures are in the marketing. The third is the one a practice owner would want. The AI Alignment Filter™ cannot evaluate a claim you have only seen the flattering half of, and the habit that protects you is simple: ask for all the numbers the study produced, not the ones selected for you.
Introduction
I want to start somewhere that is not dentistry, because the lesson lands harder from outside.
The Associated Press documented a facility outside Pittsburgh selling access to what it called medbeds. Three hundred dollars a night. Devices priced between roughly six hundred and eleven thousand dollars. Bottled water at a hundred and fifty dollars. The operation drew a warning letter from the Food and Drug Administration in 2023, and customers afterward said they had been duped.
Those people were not foolish. They were unwell, or frightened for someone who was, and they were shown confident claims by people who sounded certain.
I raise it because the distance between that and a dental exhibit hall is shorter than we like to think. Not in the honesty of the people involved, which is usually completely different. In the mechanism. Confident claims, no way to check, and a buyer who has no practical means of evaluating what they are being told.
The difference between real technology and that facility is not that real technology always works. It is that real technology is built so that when it fails, somebody finds out and publishes it.
This is a week about how to be the person who finds out.
Key Takeaways
• Funding does not make research wrong. It shapes what gets measured and which results get promoted.
• A randomized trial of AI orthodontic triage reported sensitivity of 1.0, accuracy of 0.98, and specificity of 0.67. Only two of those travel.
• One in three unnecessary referrals is the figure a practice owner needs and the one least likely to be shown.
• A sepsis model deployed at hundreds of hospitals was later found on external validation to be a poor predictor. Wide adoption is not evidence.
• Ask what the study measured, what it did not measure, and which of the reported numbers is missing from the slide.
• What makes real medicine credible is not that it succeeds. It is that its failures are detectable and published.
Three Numbers, Two of Which You Will Be Shown
Here is the cleanest teaching example I have found in dental research, and it is a good study.
A randomized controlled trial in a publicly funded orthodontic clinic examined AI-supported remote triage against face-to-face screening. It reported sensitivity of 1.0, meaning the system caught every patient who genuinely needed referral. Overall diagnostic accuracy came in at 0.98.
Those are the two numbers you will see.
The study also reported specificity of 0.67. That means roughly one in three patients who did not need a referral was flagged as needing one.
Now think about which of those three figures matters most to a practice. Sensitivity protects the patient, and it should be high. But specificity is what determines your Tuesday. A system that refers a third of the people who do not need referring generates appointments, generates parent anxiety, and generates work, and it does so invisibly because every one of those appointments looks like appropriate caution.
There is a further detail worth having. Sensitivity of 1.0 is specific to a threshold. At a stricter referral threshold, the same study reports sensitivity falling to 0.93. Move the dial and the flagship number moves with it.
And the artificial intelligence in that study is doing less than the phrase suggests. The system supplied the measurements. Orthodontic residents assigned the grades.
None of this makes the trial bad. It makes it a normal, careful piece of research that reported everything it found, which is exactly what you want. The problem arises one step later, when somebody builds a slide from it.
Report All Three, or Use a Different Study
The rule that follows is short enough to remember in a room.
If a claim rests on a study that reported three numbers, ask for all three.
This is not an accusation, and it does not require you to know which numbers a study of that type would produce. You can ask it generically: what else did that study measure, and what did it find?
A vendor who knows their evidence will answer. A vendor who has only ever seen the marketing summary will not be able to, and that is worth knowing too, because it tells you that nobody in the chain between the research and you has read the original.
Selective reporting is rarely a decision anybody made. It is what happens when a finding passes through four people, and each one keeps the part that made the point they were making.
Wide Adoption Is Not Evidence
There is a substitute for evidence that circulates constantly in health technology, and it is the observation that a great many places are already using it.
A sepsis prediction model built into one of the most widely used hospital record systems was deployed at hundreds of hospitals. When researchers eventually validated it externally, at a health system that had not built it, they found it performed poorly and missed a substantial majority of the sepsis patients it was meant to catch, while generating a heavy burden of alerts.
Hundreds of hospitals had adopted it. The adoption told you about the distribution agreement, not about the model.
I raise this because "everyone is using it" is the most persuasive non-argument in technology purchasing, and in dentistry it usually arrives as a practice count. A vendor reporting deployment in tens of thousands of practices is reporting a sales figure. It is a real achievement, and it is not a clinical finding.
The two can point in opposite directions, and when they do, the sales figure is the one that grows faster.
Who Funded It, and Why It Matters
Now the question in the title, asked properly.
Industry funding does not make a study false, and treating it as though it does would eliminate most of the research that exists in any applied field. Companies fund research because they are the ones with a reason to.
What funding reliably shapes are three things.
What gets measured. A company measures the outcomes its product is good at. A manufacturer studying procedure time is not hiding accuracy data; often nobody collected any.
What gets published. Studies that go well are far more likely to be written up and promoted than studies that go badly. This is true of academic research too, and it is worse where the sponsor controls the release.
Who is in the room? When authors hold a financial interest in the product, disclosure is required, and reading the disclosure statement takes twenty seconds. It is the most underused twenty seconds in evaluating any health technology claim.
None of that requires cynicism. It requires reading the funding line, which is printed at the bottom of nearly every paper, and asking whether the outcomes that were measured are the outcomes you care about.
The Dental Version of the Question
For your own purchasing, three practical translations.
When a clearance figure is quoted, ask what it was compared against. Most dental imaging AI is validated against a consensus of expert readers rather than against biopsy or clinical outcome. That is a reasonable standard, and it is not the same as truth, and knowing which one you are being offered changes what the number means.
When efficiency is claimed, ask what was not measured. A robotic system reporting a substantial reduction in procedure time, with no accompanying deviation or accuracy data, has told you about speed. For implant placement, accuracy is the clinical question, and its absence is itself informative.
When a practice count is offered, ask for an outcome. How many practices use it is a distribution fact. What changed in those practices is a clinical fact. The second is much harder to produce and much more relevant.
What Makes Something Credible
I want to end where I started, because the contrast is the point of the whole week.
The facility outside Pittsburgh had no mechanism for being wrong. There was no study that could have come back badly, no external validation that might have deflated it, no disclosure line to read, no regulatory file to look up. It could only ever be right, and a claim that can only ever be right is not a strong claim. It is an unfalsifiable one.
Real medical technology is the opposite, and it is less flattering. Effect sizes shrink when somebody measures them properly. External validations come back worse than internal ones. Authors describe their own results as modest. Regulators publish summaries that occasionally reveal there was no performance data at all.
That is not a field in trouble. That is a field where the checking works, and it is the only reason any of it can be trusted.
So when a technology's literature contains disappointments, take it as reassurance. The ones with no disappointments anywhere in their record are not better. They are unexamined.
Frequently Asked Questions
Is industry-funded research worth reading?
Yes, and most applied dental research is industry-funded to some degree. Read the funding line, read what outcomes were chosen, and weight it accordingly. Dismissing it wholesale would leave you with almost nothing.
How do I find the disclosure statement?
It is usually at the end of a published paper, under conflicts of interest or funding. For a product claim, ask the vendor which publication it comes from and look it up. If there is no publication, you have your answer about what kind of claim it is.
What if a study reports only favorable numbers?
Ask what else was measured. Sometimes the honest answer is that nothing else was, which is itself a finding about how thoroughly the product has been examined.
Should I avoid products without independent validation?
No. Almost nothing in dental practice management has it, and a rule that strict would stop you from adopting anything. Buy on promise where the stakes are low and reversible. Require evidence where the stakes are clinical, and the error would be invisible.
Is there a quick version of all this?
Three questions. How many, and where. What else did the study measure? And who paid for it? Ninety seconds, and it catches most of what matters.
Final Thoughts
Dentists are asked to evaluate claims constantly, from people who are more practiced at making them than we are at assessing them. That asymmetry is real, and it is not going away.
What closes it is not expertise in artificial intelligence. It is three or four questions asked every time, in the same order, until they are automatic.
The people in Pennsylvania who paid three hundred dollars a night were not given the chance to ask. You are, every time, and the questions are free.
Ask who measured it. Ask what else they found. And ask who paid.