Dental AI, Joanna Stern, Pearl, and the Risk of Overtreatment
Joanna Stern spent a year letting artificial intelligence into nearly every corner of her life. The longtime Wall Street Journal technology journalist used AI to help manage work, health care, transportation, parenting, household chores, and even a simulated romantic relationship. Her 2026 book, I Am Not a Robot, became a New York Times bestseller because it asked a practical question rather than a futuristic one. What happens when AI stops being something we discuss and starts making suggestions inside ordinary life?
For dentists, the most important part of Stern’s experiment happened in the operatory. At a dental visit using Pearl AI, colored overlays on radiographs helped identify findings that became part of a recommendation for extensive periodontal treatment. Stern was skeptical, sought several additional opinions, and ultimately did not proceed with the proposed therapy. She later reported concerns that some dental practices may be using AI findings to encourage more treatment. Her story spread quickly because it touched one of dentistry’s oldest trust problems. The person who diagnoses the disease may also be the person paid to treat it.
Pearl CEO Ophir Tanz pushed back. His argument was not that Stern imagined the disagreement, but that the disagreement was about treatment, not whether something was visible. That distinction matters. Pearl’s FDA clearance describes Second Opinion as computer aided detection software for bitewing and periapical radiographs. Its role is to assist the dentist as a second reader. The original clearance included caries, restoration margin discrepancies, calculus, and periapical radiolucencies. It did not clear the software to diagnose periodontitis, prescribe scaling and root planing, or decide which tooth should be restored.
That is where the debate becomes useful for every dental office. AI detection, clinical diagnosis, and treatment planning are not the same act. A colored box can be correct and the treatment plan can still be too aggressive. The dentist can disagree with the software and still be wrong. A patient can have disease without symptoms, and a finding can be real without requiring immediate irreversible treatment.
The periodontal example makes the point clearly. The American Academy of Periodontology staging and grading framework does not diagnose periodontitis from an isolated radiographic finding. Clinical attachment loss, probing depth, radiographic bone loss, tooth loss, furcation involvement, mobility, distribution, progression, smoking, diabetes, and other factors shape the diagnosis. Radiographic calculus may be useful information, but it is not a treatment plan. If an AI overlay becomes the dominant reason a patient is told she needs four visits of periodontal therapy, the workflow has skipped several clinical steps.
Caries presents the same problem in a different form. Modern AI can detect radiographic lesions impressively well, but early detection does not automatically mean restoration. A 2026 Clinical Oral Investigations study independently tested Diagnocat on 100 bitewings containing 1,540 approximal surfaces. Overall accuracy was 91.6 percent and specificity was 94.3 percent. Yet sensitivity was 73.1 percent, and positive predictive value was 64.7 percent for any approximal caries. When researchers asked the system to classify lesion depth, positive predictive value fell to 42 percent for enamel lesions and 39.2 percent for dentin lesions in that dataset.
Those numbers do not mean dental AI is unreliable. They mean the colored box is the beginning of a conversation, not the end of one. Positive predictive value changes with disease prevalence, patient population, software thresholds, and study design. An early enamel lesion may call for fluoride, diet change, surveillance, improved hygiene, resin infiltration, or another nonoperative strategy. Even a correct detection does not answer the question that matters most to the patient. Do we need to drill this tooth today?
The larger literature points in both directions, which is exactly why simple headlines are dangerous. A 2025 systematic review and meta analysis of AI caries detection found generally strong performance, with reported accuracy ranging from 73.3 percent to 98.8 percent across included studies. But the studies varied enormously in algorithms, radiographic modalities, patient populations, reference standards, and methods. The authors also warned that false positives can produce unwarranted intervention and stressed the importance of specificity, especially when disease prevalence is low.
At the same time, AI does not inherently create overtreatment. A 2025 randomized controlled trial involving 30 dentists and 50 panoramic radiographs found that AI assistance improved overall accuracy in diagnosing periapical radiolucencies from 91.6 percent to 93.3 percent. More importantly, false positives fell from 746 to 345. Specificity increased from 95.7 percent to 98.0 percent, and dentists showed a trend toward more conservative treatment choices. In that controlled setting, AI acted less like a salesperson and more like a brake on human overdiagnosis.
That result should temper both sides of the argument. It is too easy to say that AI will make dentists overtreat. It is equally easy to say that better detection automatically means better care. AI is more accurately understood as an amplifier. It can amplify detection, consistency, patient understanding, and clinician confidence. It can also amplify whatever incentives surround the clinical environment. In a practice culture built around careful diagnosis and conservative treatment, AI may reduce missed disease and unnecessary intervention. In a culture that measures doctors by how many AI findings convert into procedures, the same technology can become a production tool.
This is why Dr. Owais Farooqi’s call for independent local validation deserves attention. An AI system may perform well in a manufacturer’s study and still perform differently with another sensor, another patient population, another disease prevalence, or another workflow. FDA clearance establishes that a product can legally be marketed for a defined use. It does not guarantee identical performance in every practice. Dentists, groups, schools, and health systems should increasingly ask a second question. How does this tool perform on our images and our patients?
The practical lesson is not to fear dental AI. It is to build a better workflow around it. The dentist should read the radiograph first when practical, then use AI as a second reader. The AI finding should be reconciled with the clinical examination, history, risk assessment, probing, symptoms, lesion activity, and longitudinal records. When the recommendation is irreversible, the threshold for confirmation should rise. Patients should be told what the software sees, but also what remains uncertain and what options exist besides immediate treatment.
That approach may actually improve case acceptance because trust grows when patients can see both the finding and the reasoning. The worst use of AI is to put a red box on the screen and say, the computer says you need this. The better use is to say, the software highlighted this area, I agree that something is present, and here is why I believe we should treat it now, watch it, or manage it preventively. The technology becomes evidence rather than authority.
Stern’s story matters because it exposes a new form of an old dental problem. Dentistry has always balanced the risk of missing disease against the risk of treating too much. AI does not remove that tension. It makes the evidence more visible, more persuasive, and sometimes more difficult for patients to question. The profession’s responsibility is to make sure that stronger detection does not become automatic intervention.
Pearl was cleared to help dentists see more. The dentist still has to decide what those findings mean for the person in the chair.
When AI puts the colored box on the X ray, what happens next in your practice?
Join the Conversation!