Does AI caries detection work differently for novices and experts — a randomized trial measuring the effect on the treatment threshold — a clinical trial (ClinicalTrials.gov)

1From finding it to deciding whether to drill

Proximal caries, which form where teeth touch, are hard to see directly and are found through radiographs. But a shadow on an image does not automatically mean drilling. Early demineralization is often handled with fluoride application or watchful waiting, and where to set the point of switching to restorative treatment is left to the reader's judgment.

Introduce an AI that helps with detection and the effect does not stop at detection. If more lesions are found, the number of occasions demanding a decision about drilling rises too. This study takes that chain head-on.

2The hypothesis that the user changes the effect

Hand out the same AI, and the result can differ with the recipient's level of training. Making that comparison the focus is somewhat unusual for an evaluation of medical AI. Most trials measure model accuracy or patient outcomes; the question here is how AI changes human judgment — and it starts from the premise that the change is not uniform.

The contrasts set out above are the hypotheses the study means to test, not results it has shown.

3Assigning readings, not procedures

The intervention is the review of radiograph images, not the assignment of different procedures to patients. It is classified as interventional, but what is actually manipulated is the condition of the reader. Detecting caries is among the most frequent judgments made in dentistry, and this design measures what happens when AI enters that judgment without exposing patients to treatment.

Of the 803 AI-related clinical trials this site holds as of 2026-09-02, 25 relate to dentistry or caries — not a large share.

4A question placed in the setting of education

Early readers are at the stage of building their own standards. What it does to the formation of those standards to receive AI prompts every day is a question separate from accuracy. Following the AI when it is right may be desirable, but whether the ability to suspect a lesion the AI did not flag still develops is unknown.

A target of 60 keeps this to early validation, yet the question it raises is not confined to dentistry: it appears wherever imaging AI enters a training setting.