WSJ columnist Joanna Stern's book I Am Not a Robot, discussed on the New York Times' Hard Fork podcast May 8, surfaces a problem with dental AI screening tools that builders selling into clinical workflows should track. Pearl AI flagged Stern's dental scan as needing four sessions of periodontal treatment costing thousands; multiple dentists reviewing the same images disagreed. Pearl AI and Overjet are the named vendors, both well-funded incumbents in the dental-radiograph AI space. Pearl's marketing pitch to dentists, quoted in the Futurism writeup: practices using it find "37 percent more disease and deliver 24 percent more care to patients."
The marketing claim is the architecturally interesting part, separate from any one patient's experience. An AI product sold to its operator on revenue uplift โ "you'll bill more after deploying us" โ is qualitatively different from one sold on accuracy improvement โ "your diagnoses will be more correct." Pearl's pitch combines both: the 37% surfaces as a detection-improvement claim, the 24% as the billing consequence. If the deployment data supports the detection number on out-of-sample patient cohorts compared against expert consensus, the billing follows. If the detection number is calibration drift toward false positives, the billing is iatrogenic. The article doesn't surface published validation studies, FDA-clearance-grade comparative reads, or third-party audit data that would let an outside observer tell which.
Place this against this morning's Ontario auditor-general audit: 20 of 20 AI scribe vendors showed inaccuracies during procurement testing. Two domains, two failure modes โ note-fabrication on the scribe side, possible diagnostic-inflation on the imaging side โ but the same underlying gap: deployment is outrunning third-party evaluation. Dental AI vendors don't yet face the kind of provincial-audit scrutiny Ontario applied to scribes; the US dental market is private-payer-dominated, which changes the regulator's leverage. Watch the FDA, state dental boards, and major dental-insurance carriers as the bodies that could publish their own comparative studies once enough Stern-grade anecdotes accumulate.
Monday: if you build or sell clinical AI, the line between "improves detection" and "increases billable procedures" is the line your marketing copy should not blur, regardless of how the dental market currently rewards the blur. If you're a dentist deploying these tools, the audit-grade question is not what the vendor's claimed improvement number is โ it's whether a second-reader human dentist agrees with the AI flags at a rate that exceeds inter-rater human agreement on the same images without AI. That's the harness; the vendor either has the data or doesn't. If you're picking which clinical AI segments to enter, the absence of public validation data in dental imaging is either an opportunity (be the vendor that publishes it) or a warning (the segment may be repriced when someone else does).