In 2021-22, I worked as a dental X-ray annotator for an AI healthcare company. (one of the early humans behind the scenes teaching machines to read dental radiographs)
The work was painstaking and precise. OPGs, bitewings, intraoral periapical X-rays, across all of them, my job was to mark exact anatomical boundaries: where enamel ended and dentin began, where the PDL space was, where bone levels sat, where pathology existed. Every marking I made became training data. Every correction I submitted shaped what the model learned next.
Here’s what struck me most about that period:
The machine wasn’t bad at dramatic pathology.
Large cysts, obvious abscesses, clearly missing teeth, it handled those reasonably well. What it consistently struggled with was subtlety. Early-stage bone loss. Ambiguous carious lesions. Distinguishing an existing restoration from new decay. The exact things a clinician develops an eye for over years of looking.
It couldn’t read context.
A human annotator looks at an X-ray and unconsciously factors in dozens of things simultaneously like patient age, tooth position, surrounding bone pattern, likely clinical history. The model in 2021 was reading pixels. It didn’t know that a shadow in one location means something completely different from a shadow in another.
The errors were confident.
This was perhaps the most unsettling part. The model didn’t flag uncertainty but it misidentified bone levels or misread restorations with the same output weight as a correct diagnosis. That kind of confident incorrectness is clinically dangerous in a way that obvious uncertainty isn’t.
What changed my perspective wasn’t how wrong the machine was then but how rapidly it improved once enough clinical minds had contributed enough precise annotations. The progress between 2021 and where dental AI sits now is genuinely remarkable.
But it also confirmed something I’ve held onto since: AI in diagnostics is only as trustworthy as the clinical expertise that trained it. The model doesn’t teach itself. It learns from annotators who know what they’re looking at.
Which means the future of AI in dentistry isn’t less clinical judgment but it’s clinical judgment applied upstream, at the point of training, not just at the point of treatment.
Do you think AI will eventually match, or maybe exceed, the diagnostic accuracy of an experienced dental clinician? And does that prospect excite or concern you?
MBH/DB