When AI flags every caries on the radiograph, does the dentist's job shift from diagnosis to defending against over-treatment?

The shift

Detecting a lesion on a radiograph goes from scarce (a trained eye that catches what others miss) to abundant: an AI reads the film in seconds and flags every caries, including the sub-clinical and borderline ones a human would have passed over. Pattern-matching against the image — the thing dental training spent years building — is now something Overjet or Pearl does by default, at scale, on every X-ray. The scarce act inverts: not spotting the lesion, but deciding which flagged lesions deserve a drill and owning that call.

The axioms

• Finding pathology in an image is gated by scarce, trained visual perception.
• The exam is detection — the dentist's differentiator is seeing the decay a lesser clinician misses.
• Finding a lesion implies treating it; a flagged caries is a caries to be fixed.
• Treatment is a physical act on a physical tooth — the drill, the fill, the crown, the hand.
• A licensed, accountable clinician owns whether to treat or watch.
• The patient trusts the dentist's read because they can't interpret the film themselves — the asymmetry justifies the recommendation.
• Borderline lesions are judgment calls made under uncertainty about progression, risk, and the specific patient.

Invalid axioms

1. Finding pathology in an image is gated by scarce trained perception. The AI flags everything, including incipient lesions below the threshold a human eye reliably catches. The habit-trap: practices still market and price "a thorough exam" around the dentist's diagnostic acuity, when detection is now the commodity input and the scarce work sits entirely downstream of it.
2. The exam is detection, and catching more is being a better dentist. Sensitivity is no longer the differentiator — the tool out-detects the human, and its bias runs toward over-flagging borderline demineralization. The habit-trap: "we found something you'd have missed elsewhere" still reads as a mark of quality, when a rising find-rate now signals a calibration problem as easily as diligence.
3. Finding a lesion implies treating it. This reflex was safe when detection was hard: if a trained eye saw it, it was usually real and advanced enough to act on. Once every arrested, remineralizing, or radiographically-ambiguous spot gets flagged, "flagged" and "should be drilled" come apart. The habit-trap: charting, treatment planning, and — critically — insurer and DSO incentive structures still treat a positive finding as a green light to bill and cut, converting AI sensitivity directly into intervention.

Unchanged axioms

1. Treatment is a physical act on a physical tooth. Removing decay and placing a restoration is irreversible work by trained hands — no amount of detection abundance touches this. And because it's irreversible, a wrong "treat" call has an asymmetric cost the AI never bears.
2. A licensed, accountable clinician owns the treat/no-treat call. Liability doesn't transfer to Overjet. When a healthy tooth gets drilled on the strength of an AI flag, a named person with a license answers for it — not the model, and not cleanly the insurer or DSO that pressured the call. This is a legal fact, so it holds even as the tools improve; if anything the accountability sharpens as detection stops being the clinician's contribution.
3. Borderline lesions are judgment under real patient-specific uncertainty. Whether an incipient lesion warrants a filling, a fluoride varnish, or watchful waiting depends on caries risk, diet, hygiene, prior progression, and the patient in the chair — exactly the novel, low-pattern context where matching against "every radiograph ever labeled" is weakest. Deciding not to treat, and being able to defend that restraint, is now the scarce skill. This is the sharpest inversion in the audit.
4. Patient trust is the thing that survives the asymmetry flipping. The old asymmetry (the patient couldn't read the film) is eroding — the AI overlay is shown to patients as a persuasion tool, and patients can increasingly get a second read. What can't be manufactured is a clinician the patient believes is choosing in their interest rather than the practice's. Trust stops being a byproduct of expertise and becomes the explicit thing being spent or protected.

New axioms

1. A flag-rate that rises with detection sensitivity manufactures over-diagnosis, and someone monetizes it. When the tool surfaces every borderline lesion and the business model rewards intervention, AI sensitivity converts into drilling by default. We must solve for how a positive finding gets adjudicated before it becomes a treatment plan — the flag is now the start of a decision, not the decision.
2. Insurers and DSOs can weaponize the AI read in both directions. The same flag that a DSO uses to justify a filling, an insurer can use to deny a claim ("the AI shows it's watchable") or to audit a dentist who treats "too little." The clinician is squeezed between a tool that over-detects and two payers reading its output to their own advantage. We must solve for who governs how AI radiograph output is allowed to drive reimbursement and treatment authorization.
3. Who protects the patient from AI-justified over-treatment? The patient now faces a recommendation backed by an authoritative-looking overlay, made by a clinician who may be incentivized or pressured to treat. Nothing in the current structure is designed to defend the no-treat call — there's no scarce, funded role whose job is restraint. We must solve for where the check on over-intervention lives when detection is free and intervention is profitable.
4. The evidence base for treating incipient lesions was built in a low-detection world. Clinical guidance on when to fill vs. watch was calibrated to lesions humans actually saw. Feeding it a flood of sub-clinical flags applies old thresholds to a new detection floor, and it's unclear the thresholds still mean what they meant. This one is moving fast — as detection models get more sensitive, the mismatch widens unless the treat/watch thresholds are re-grounded.

Where it breaks

"The exam is detection, and finding more is doing better dentistry" (invalid) collides with "a licensed clinician owns the treat/no-treat call and answers for it" (still holds). Once the AI out-detects the human, the dentist's remaining contribution is judgment and restraint — the decision not to drill a flagged-but-stable lesion. But the incentive structure around them (DSO production targets, fee-for-service billing, the overlay shown to the patient as justification) still rewards acting on findings, and the liability still lands on the clinician who acts. The dentist is being asked to own restraint while every force around them — the tool's sensitivity, the employer's targets, the persuasive overlay — pushes toward intervention.

A second collision: "finding a lesion implies treating it" (invalid) meets "who protects the patient from AI-justified over-treatment" (new). The reflex that was safe when detection was scarce is now the exact mechanism by which over-treatment scales — and no one has been assigned the job of interrupting it. The flag flows straight to the treatment plan, and the only thing standing between it and the drill is a clinician who has to spend their own trust and accept their own liability to say no.

ai-axioms.vercel.app/axioms/healthcare/dentistry-overtreatment-ai-axioms