What Is the AI in Your Operatory Actually Cleared to Do?

You can answer this in about a minute, and almost nobody does. The FDA maintains a public, searchable record of every cleared device, including what each one is permitted to claim. In dentistry, one pillar of AI is regulated, and the rest is not. Radiographic AI, the software reading your bitewings and panoramics for caries, periapical radiolucencies, calculus, and bone levels, is cleared as a Class II medical device. Crown design, voice charting, scheduling, caries-risk prediction, and patient chatbots are largely unregulated software. A peer-reviewed review of the FDA record identified 29 cleared dental imaging modules across 13 companies, and every one of them is radiographic. The AI Alignment Filter™ works on both categories. What changes is how much checking is already done, and how much is yours.

Introduction

There is a search box on the FDA's website. You type the name of a product into it and, if the product has been cleared, you get the clearance number, the decision date, and a summary document stating what the device is intended to do.

It takes a minute. I would guess that fewer than one dentist in twenty has ever done it for a product in their own operatory.

That is not a criticism of dentists. It is a statement about how technology is presented to us. The clearance appears in the marketing material as a badge, not as a document, and a badge is designed to end the inquiry rather than start it.

But the document underneath the badge is where the useful information lives. It tells you what the thing was tested on, what it is allowed to say, and sometimes it tells you that the performance data supporting it was thinner than the badge implies.

This week is about the difference between what is cleared, what is proven, and what is simply sold. In dentistry, those three categories do not overlap as much as the exhibit hall suggests.

Key Takeaways

•     Radiographic AI is the only clinically regulated pillar of dental AI, and it has reached real scale.

•     Crown design, scribing, scheduling, risk prediction, and chatbots are largely unregulated software, which is legal, common, and not an accusation.

•     Clearance means a claim was supported well enough to permit sale. It is a floor, not a verdict, and one cleared product's FDA summary contains no performance figures at all.

•     Two credible counts of cleared dental AI disagree because they are counting different things. Knowing which unit is being counted is most of the skill.

•     Model accuracy drops when a model meets a population it was not trained on, and this is normal rather than scandalous.

•     FDA decision dates are public. Press release dates are not the same as the decision dates.

One Pillar Is Regulated. The Rest Is Software.

Dental radiographic AI is computer-assisted detection software. It runs on bitewings, periapicals, panoramics, and increasingly cone-beam CT, and it marks suspected caries, periapical radiolucencies, and calculus, and measures interproximal bone levels in millimeters.

It is regulated as a Class II medical device because it makes a claim about pathology in a patient. That is the trigger. When software tells a clinician something about disease, it enters the regulatory system.

Now look at the rest of what is marketed to a practice as artificial intelligence.

Crown and restoration design software takes a scan and proposes a finished restoration. It is design and manufacturing software, not diagnostic software, so it carries no diagnostic clearance, and none is required.

Voice-driven periodontal charting and AI scribing are not regulated as medical devices. There are no FDA clearances for the voice-charting products in this market, and every time-saving figure attached to them is reported by the vendor or by a customer the vendor selected.

Front-office automation, scheduling, reactivation, and patient chatbots sit outside the device framework entirely.

Caries-risk prediction is the interesting case. It sounds clinical, it is often presented clinically, and no risk-stratification model appears among the FDA-cleared dental AI products at all.

None of this makes the unregulated tools bad. Some of them are the best value in the building. It means that for one category somebody independent has examined evidence and published a document, and for the other category the checking is entirely yours.

What a Clearance Actually Says

Clearance is narrower than it sounds, and reading one is a genuinely useful skill.

It states an intended use. Not a capability, an intended use, and those differ. A product cleared to detect caries on bitewings has not been cleared to detect anything else, on any other image type, even if the software will happily draw a box around it.

The performance figures supporting dental imaging clearances cluster in a recognizable band: sensitivity in the high eighties to low nineties, specificity in the high seventies to high eighties, area under the curve generally between 0.85 and 0.95. Those are established in retrospective reader studies, against a ground-truth set by consensus among board-certified dental radiologists or oral surgeons.

Read that last sentence again, because it is the honest limitation. The comparator is not biopsy or clinical outcome. It is what a panel of experienced readers agreed they saw. That is a reasonable standard, and it is the standard available. It is not the same thing as truth.

And clearance does not guarantee that any performance figure exists at all. In the peer-reviewed review of cleared dental AI, one cleared cephalometric product's FDA summary was found to provide no quantitative performance metrics whatsoever. The product is legitimately cleared. There is simply no number in the public file to read.

The strongest evidence base in this space belongs to a product most general practices do not own. DentalMonitoring, which uses smartphone-captured intraoral images to monitor orthodontic treatment remotely, is the only dental AI product authorized through the De Novo pathway in May 2024, and it carries fourteen peer-reviewed publications behind it. That is what a deep evidence base looks like in this field, and it is rare.

The Best-Sourced Number in Dental AI

If you want to see what a good number looks like, here is one.

Overjet's bone-level measurement differs from a consensus ground truth by roughly three tenths of a millimeter. Specifically, 0.307 mm on bitewings and 0.353 mm on periapicals, adjudicated by dental radiologists.

Three things make that figure better than almost anything else in this market.

It comes from the FDA clearance summary rather than from a brochure. It is a measurement rather than an outcome, so it is not entangled with how a clinician responded to it. And it reports to three decimals across two image types, which is what a real measurement looks like as opposed to a round number chosen for a slide.

Notice what it does not claim. It does not say patients do better. It does not say your hygiene department will produce more. It says this software measures bone level to within about a third of a millimeter of what expert readers agree it is.

That is a modest claim; it is carefully bounded, and it is exactly why you can trust it.

Two Counts That Disagree, and Both Are Right

Here is a small puzzle that teaches more than it looks like it should.

One analysis of the FDA 510(k) database identified 44 AI-enabled dental software products cleared between 2021 and 2025, with 18 of those in 2025 alone.

A peer-reviewed review published in the International Dental Journal identified 29 distinct cleared imaging modules across 13 companies.

Both are competent. They are not in conflict. They are counting different things: all cleared dental software as a medical device in one case, imaging modules only in the other.

A fifty percent difference between two counts is not evidence that someone is wrong. It is evidence that the unit of measurement was never stated. This happens constantly in technology claims, and once you start looking for the unit you will find that a surprising number of impressive figures do not have one.

When a number is quoted at you, the first question is not whether it is true. It is what is being counted.

How a Company Counts Its Own Clearances

That principle has an immediate and practical application.

Pearl, one of the significant vendors in this space, states that it holds 26 FDA clearances. The FDA database shows the company holding eight dental 510(k) clearances. The peer-reviewed review credits Pearl with seven cleared modules.

Three numbers, none of them dishonest. The 26 counts individual cleared capabilities across the clearances. The eight counts clearance documents. The seven counts modules as the reviewers defined a module.

I am not raising this to embarrass a vendor, and Pearl's cleared products are real. I am raising it because a regulatory-literate reader will challenge the 26, and because you are going to be shown a number like it by somebody this year.

The right response in the room is not skepticism. It is a question with no edge on it: what is being counted there? Any vendor who knows their own regulatory file answers that easily.

The Number That Falls When It Leaves the Lab

One more pattern, and it is the one that generalizes furthest.

A model predicting periodontal treatment response, published by Feher and colleagues in the Journal of Periodontology, reached an area under the curve of 0.93 on the data it was developed on. On external validation, in a different population, it fell to 0.76.

That drop is the normal behavior of predictive models, not a scandal. A model learns the population it was trained on, including that population's particular mix of disease severity, imaging equipment and recording habits. Show it a different population and some of what it learned stops applying.

There is a further wrinkle worth knowing, because it cuts the other way. The external cohort in that study had milder disease, which means the fall in performance mixes two separate effects: the model meeting an unfamiliar population, and the model facing a harder discrimination task. Those are different problems, and the single number cannot separate them.

The practical translation for a practice owner is short. A performance figure produced on the developer's data is an upper bound. Your patients are the external cohort.

Reading the Date

Small thing, quick to check, and it tells you something about how a claim reached you.

Three FDA clearance dates circulating in dental AI coverage are press release dates rather than decision dates. The FDA record shows Overjet's CBCT Assist cleared on 5 December 2025, VideaHealth's Caries Assist on 21 April 2022, and the Neocis Yomi S on 28 October 2025. Each of those differs from a commonly cited date by one to three weeks.

Nothing sinister is happening. A company announces when its communications team is ready, and the announcement date becomes the date everyone repeats.

But it is a free, one-minute test of whether whoever is talking to you went to the primary record or to the press release. Ask which 510(k) number a product was cleared under. It is a friendly question; it has a correct answer, and the answer is public.

Running the Filter Chairside

The AI Alignment Filter™ runs the same four steps in an operatory as it does anywhere else, but the clinical setting sharpens the fourth one considerably.

Name the bottleneck first, in a sentence. Not "we should be using AI." Something like: interproximal caries on bitewings are being called inconsistently between my two associates.

Then the cognitive load. Is this tool helping someone think, create, organize, or capture? Voice charting removes a capture burden and a second person from the operatory. Radiographic AI addresses a thinking burden, and those are not interchangeable purchases.

Then integration cost, which is where dental adoptions die, exactly as they do everywhere else. The subscription is never the cost. The cost is the eight weeks where the team runs the new way and the old way at once, plus the person who has to own it on top of a full schedule.

And then the hard gate. Is this replacing judgment or enhancing it?

In practice, that question has a bright line under it, and it does not move. AI may assist with reminders, reactivation, confirmations, draft communication, documentation, templates, and scheduling drag. It does not touch clinical decision-making or treatment planning.

A tool that marks a suspicious area on a radiograph for you to examine is enhancing judgment. A tool whose output reaches a treatment plan without a clinician having genuinely evaluated it has replaced judgment, whatever the interface says, and it fails the Filter no matter how well it performed on the first three steps.

Cleared is not the same as proven. Proven is not the same as proven here.

Frequently Asked Questions

Does a 510(k) clearance mean the FDA tested the product?

No. The 510(k) pathway establishes that a device is substantially equivalent to one already legally marketed. The manufacturer submits performance data, and the FDA reviews it. It is a real review of real evidence, and it is not an independent trial run by the agency.

Is unregulated dental software a red flag?

No. Crown design software and scheduling tools are not making claims about pathology, so they are not devices, and there is no reason they should be regulated. The flag goes up only when unregulated software is presented as though it carries clinical authority.

Should I ask a vendor for their 510(k) number?

Yes, and it should be an easy conversation. The number is public, the company knows it, and it lets you read the intended use statement yourself rather than the version on the booth panel.

Why is there no AI for oral cancer screening?

There is research, some of it promising, and among the cleared dental AI products in the peer-reviewed review, there is no soft-tissue oral-cancer detection product. Every cleared product is radiographic. That gap is worth knowing before a device is marketed to you as filling it.

My associate wants a tool I think fails the Filter. How do I handle that?

Run the four steps together rather than delivering a verdict. Step one usually settles it, because the disagreement is normally about which bottleneck matters rather than about the tool. If you both name the constraint in a sentence and the sentences differ, that is the actual conversation.

Final Thoughts

The dentists I talk to are not afraid of this technology. That is worth saying plainly, because the trade press keeps describing a resistance that I do not think exists.

What they are is tired. Tired of buying things that made Tuesday harder, tired of training on systems that were replaced before the training finished, and tired of being told that the thing they just bought is now obsolete.

The defense against that is not caution. It is a question asked in the right order, before the money moves, and a willingness to look up the primary record rather than accept the badge.

Cleared is not the same as proven. And if it does not replace or compress something, it may be unnecessary.