Showing AI Radiograph Overlays to Patients Without Overselling
The overlay works. That’s the problem.
Practices running Pearl’s Second Opinion or Overjet Dental Assist in the surgery report the same thing: turn the screen, show the coloured boxes sitting on the bitewing, and the conversation shortens. Patients who spent years squinting politely at greyscale films suddenly see something. Pearl’s own customer data claims a 30% lift in treatment plan acceptance; Overjet has published case studies with DSOs showing restorative case acceptance moving from roughly 40% to the high 50s. Independent UK numbers are thinner, but principals I’ve spoken to who run overlays chairside describe a real shift, particularly on borderline interproximal lesions that previously ended in “let’s watch it.”
That lift is exactly why ai radiograph patient communication needs a script before it needs a bigger monitor. A coloured box on a bitewing is a probability estimate rendered as a geometric certainty. The patient does not see a probability. They see a red rectangle labelled caries and they hear a machine agreeing with you. When the restoration turns out to be more conservative than the overlay implied, or when the “caries” was a cervical burnout artefact and you drilled anyway, the consent record has to show that the patient understood what the overlay was and was not. Right now, in most practices, it doesn’t.
What Montgomery actually requires of you here
The UK consent standard is Montgomery v Lanarkshire Health Board (2015): you must disclose material risks and reasonable alternatives, where materiality is judged by what this particular patient would attach significance to. The GDC’s Standards for the Dental Team Principle 3 builds on that with the requirement that patients give valid consent based on information they can understand.
Now apply that to a screen showing an AI overlay. The patient has been given a visual claim. If that claim is stated or implied as fact and it’s actually a 0.62-confidence detection on a lesion that may be E2 rather than D1, the patient has consented to treatment on a premise you knew to be softer than presented. That isn’t a hypothetical audit failure. It’s the substance of a complaint to the practice, then to the Ombudsman, then possibly to the GDC, and the overlay screenshot will be in the notes as evidence of what you showed them.
Add to this the DCB0129/DCB0160 clinical safety standards in NHS contexts, the MHRA’s position that these tools are Class IIa software medical devices, and the FDA’s 510(k) clearances for Pearl and Overjet which are explicitly for concurrent read and detection aid use, not autonomous diagnosis. The regulatory paperwork says “aid.” Your chairside language needs to say the same thing.
The specific sentences that create the risk
Here are the ones I hear most, and what’s wrong with each.
“The AI found four cavities.” Attributes the finding to the software. You are now a bystander to your own diagnosis, and if the count changes after you open the tooth, the patient’s trust goes to the machine, not you.
“This is 94% accurate.” Almost always a misquoted sensitivity or AUC figure from a vendor deck, applied to an individual tooth where it means nothing. Pearl’s regulatory submissions report per-condition performance, not a single global number, and the numbers vary a lot by condition. Caries detection performs differently from calculus, which performs differently from periapical radiolucency.
“The computer doesn’t miss things.” Flatly false and easily disproved by the patient’s next set of bitewings.
“It’s picked up decay you can’t see yet.” Sounds reassuring. Also describes a lesion that may be entirely appropriate to monitor under SDCEP guidance rather than restore. You have converted a monitoring candidate into a drilling candidate using the software as the argument.
Language that keeps your judgement in frame
The structural fix is simple: the overlay is evidence you interpret, never a verdict you relay. Three sentences in a fixed order do most of the work.
Sentence one, ownership. “I’ve looked at these X-rays and I want to show you what I’m seeing, with the software’s markings on top.”
Sentence two, the tool’s job. “This tool flags areas worth a second look. It’s a bit like a spellchecker. It catches things I might skim past, and it also underlines words that are perfectly fine.”
Sentence three, your read. “Two of these four marks match what I found when I examined you. This one on the upper left, I disagree with the software. And this one here I want to keep an eye on rather than treat today.”
The spellchecker line is worth stealing wholesale. It is the only analogy I’ve found that non-technical patients grasp instantly and that carries the false-positive concept without alarming anyone. Patients already know a spellchecker is useful and frequently wrong about surnames.
Notice what sentence three does. By disagreeing with the software out loud in front of the patient, on at least one finding per appointment where it’s honest to do so, you establish that your judgement is the operative thing in the room. It also makes the findings you do endorse land harder. The clinician who agrees with every box has told the patient the boxes are the diagnosis.
A worked example: the borderline interproximal
New patient, 34, two horizontal bitewings, Pearl Second Opinion running on the practice’s Dentally-integrated viewer. The overlay returns:
UR6 distal caries confidence 0.88 depth: dentine
UR5 mesial caries confidence 0.41 depth: enamel
LL7 distal caries confidence 0.79 depth: dentine
LR6 restoration (existing, flagged margin)
UL4-UL6 calculusClinically you agree on UR6 and LL7. UR5 you think is overlap from a slightly off-angle film. The calculus is real and you were going to mention it anyway.
The overselling version: “The AI’s found three areas of decay and some tartar build-up. We should get those sorted.”
The defensible version, roughly forty seconds longer:
“Right, so this software has marked five things and I want to go through them, because I don’t agree with all of it. These two, here and here, the back tooth top right and the one down here on the left, those match what I felt when I examined you. They’re into the softer layer under the enamel and I’d want to restore both. This third one, the software’s fairly unsure about it, and honestly so am I. The way this film was taken, the teeth are slightly overlapping, which can create a shadow that looks like decay and isn’t. My plan is to take one more angled film today or check it again in six months, your choice. This mark along the gumline is hardened plaque, which is accurate and which the hygienist can deal with. And this older filling has a marked edge, which I’d leave alone for now because it’s not causing you problems.”
That’s one paragraph. It has produced: two definite restorations, a documented uncertainty with two offered alternatives (Montgomery’s “reasonable alternatives” limb, discharged on the record), a hygiene referral, and a watch item. The patient has also watched you override the machine twice, which is worth more to long-term trust than the case acceptance you gained today.
Documentation, which is where practices are weakest
Showing the overlay and not recording that you showed it is the worst of both worlds. You’ve taken on the influence and none of the audit trail. Whatever PMS you’re on, Software of Excellence, Dentally, Carestream R4, a fixed note template costs nothing:
AI overlay (Pearl v2.x) shown to pt on screen.
Findings discussed: UR6 D, LL7 D - clinician-confirmed, restore.
UR5 M - AI flag, clinician disagrees, likely overlap artefact.
Options given: repeat angled BW today / review 6/12. Pt chose review.
Explained tool = detection aid, not diagnosis; clinician decision recorded above.
Pt confirmed understanding.Six lines. That template is the difference between “the dentist showed me a computer scan that said I had four cavities” and a record showing informed, clinician-led consent with the tool’s role explicitly bounded. If your practice runs overlays chairside and doesn’t have something like this in the template list, that’s the highest-value hour of admin available to you this month.
The commercial argument, honestly stated
There is a reading of all this that says caution costs money. Show the boxes, say less, book the treatment. And in the short term, yes, the vaguer version probably converts slightly better.
Two things make that a bad trade. The first is that remedial complaints are expensive in a way that dwarfs the marginal case. A single GDC investigation consumes hundreds of hours of principal time before any finding is made. The second is subtler: patients who accept treatment because a computer told them to are not loyal to you. They’re loyal to the evidence. When the next practice down the road shows them a different overlay with different boxes, you have no relationship to fall back on, because you spent the appointment being the person who read out the machine’s output.
Associates need specific attention here. If you have three associates and a hygienist all using the same overlay with different chairside patter, your consent exposure is set by whoever oversells the most. Calibration is a ten-minute item at a practice meeting: everyone runs the same three-sentence structure, everyone commits to voicing at least one disagreement with the software per session where it’s clinically honest, and everyone uses the same note template. For the wider question of which tools to run and how they perform on UK radiograph sets, our AI radiograph reading guide covers the procurement side.
One last thought on the tools themselves. Ask your vendor, in writing, for the per-condition sensitivity and specificity on a population resembling yours, and ask whether the confidence score is exposed to the clinician in the UI. Overjet and Pearl both surface some version of it; whether your particular integration displays it is a configuration question worth asking before renewal. A tool that hides its uncertainty from you makes it impossible to be honest with the patient about it, and that’s a procurement problem long before it’s a consent one.