AI can improve patient care and strengthen medical training, but only when it supports clinical reasoning instead of replacing the process that builds it.
By the time a trainee brings a patient with cancer to tumor board, AI may already have summarized the chart, organized the imaging and pathology, compared guidelines, and flagged possible trials.
That is a remarkable opportunity. Cancer care is complex, the evidence moves quickly, and clinicians lose hours to searching records. Used well, AI searches faster and catches what would be missed. It puts specialist-level knowledge within reach of clinics that never had it, and hands time back to clinicians.
But the order matters.
The learning happens before the answer
Seeing the correct recommendation has never been what builds clinical judgment. That happens earlier, in deciding which details matter, noticing what's missing, and committing to a plan while the picture is still incomplete.
Patients don't arrive as board-style questions. A scan may be unclear, and two treatments may both be guideline-supported: one offering greater disease control at the cost of more toxicity, the other fitting better with the patient's other conditions, their family, or what they want the next year to look like.
If AI performs that first synthesis every time, a trainee learns to follow a polished explanation without learning to construct one. I can follow a model's reasoning the whole way through, agree with every step, and be unable to rebuild any of it an hour later. The following felt like understanding at the time.
We call this never-skilling. De-skilling, where an existing ability fades through disuse, is documented: endoscopists who worked routinely with AI-assisted detection later found fewer adenomas once back to scoping alone. Never-skilling is quieter. The skill may never develop at all, so there's nothing to notice fading.
Why this matters for patients
A convincing answer is not enough. When researchers checked chatbot responses against cancer treatment guidelines, roughly a third contained at least one recommendation that didn't match, and some included invented information.
Catching that takes a clinician who can tell when the evidence doesn't fit and will say plainly what's uncertain. Those abilities get built in training, in the difficult cases where the answer isn't obvious: you form an assessment, defend it, find out what you missed, revise. That effort is not a slow step on the way to learning. It is the learning.
A second reader, not the first thinker
The approach we propose is simple. Reason first, use AI second, verify always.
- The Socratic pause. Before consulting a model, the trainee writes down a preliminary assessment, a plan, and the uncertainties that remain.
- AI as second reader. The model then checks the stage, surfaces an omission, or challenges an assumption.
- Provenance checks. Important claims get traced back to primary evidence, current guidelines, and this patient's actual details.
- Occasional AI-free discussions. The point isn't rejecting the technology; it's making sure clinicians can still reason when a tool is unavailable, uncertain, or wrong.
The comparison itself is the teaching moment. What did the trainee see that the model missed, and what did the model surface that the trainee overlooked? In the article we map this across the training activities where the stakes are highest: what the trainee does alone, what the model is for, and what's being assessed.
Better clinicians, better use of AI
An AI-free future is neither realistic nor desirable. What the next generation needs is two competencies: independent clinical reasoning, and safe AI-assisted practice. The strongest clinicians will know when to reach for these tools, how to question them, and how to turn what comes back into care that reflects the evidence and the patient's values.
Let AI reduce the work that drains attention, not replace the work that develops judgment. Used in that order, it can make clinical reasoning better informed, better tested, and more responsive to every patient.
Article Details:
Before the Answer: Preventing Never-Skilling in Oncology Training
Zara Baloch, Yan Leyfman, and Nikhil G. Thaker
First published: July 30, 2026
DOI: 10.1177/2993091X261474371
AI in Precision Oncology
