When we trained in radiation oncology 20–30 years ago, learning was physical. We positioned patients, marked skin with ink, and designed fields on simulators against anatomy we could see and touch.
When planning moved to CT, manual contouring forced hundreds of spatial decisions per case, building the perceptual expertise that lets an experienced clinician look at a dose-volume histogram and dose distribution and know—within seconds—that something is wrong, even when every constraint is technically met.
AI auto-contouring is faster, more consistent, and often more accurate than junior trainees. The efficiency argument is sound. But the field has not asked what that contouring was building developmentally, or what happens when the task disappears before residents acquire the cognitive architecture it was producing.
This transition is different from those before it. IMRT required more contouring. Image-guided radiotherapy required more attention to anatomical variation. The AI transition is the first to remove a core training task rather than transform it—and it removes the task doing the most developmental work.
The deeper concern is not that AI will make errors. It is what cognitive architecture is required to catch AI errors in a domain where you have never built independent judgment. An experienced physician checks AI contours against a model built from thousands of manual decisions.
A resident who has only audited AI outputs pattern-matches against a much thinner foundation—and is most likely to miss the errors.
OAR contouring can transition to audit-and-correction—but only after enough manual contouring to build spatial fluency. The point was never to produce contours. It was to build the perceptual foundation that makes supervising any contour meaningful.
AI will likely extend to GTV/CTV delineation soon. Protect GTV/CTV as a training exercise. This is not a spatial task—it is clinical reasoning expressed spatially. Integrating imaging, disease biology, and patient-specific judgment is the core intellectual work of our specialty. It cannot be developed by auditing AI outputs.
Finally, senior faculty need to make their reasoning explicit, not just visible. What do you actually see when you review a plan and immediately sense something is wrong? That implicit knowledge is exactly what residents need and cannot absorb through observation alone. Thinking out loud and discussing is probably the highest-leverage teaching act available right now.