This Resident Roundup post was submitted by Ruben Arias, MD, a third-year orthopaedic surgery resident at the Universidad Pontificia Bolivariana in Medellín, Colombia.
It was 2 a.m. on a trauma call when I caught myself doing it. A 28-year-old motorcyclist had come in with a floating shoulder, and while I was waiting for the CT, I opened ChatGPT on my phone and typed: “management of ipsilateral clavicle and glenoid neck fracture.” The answer was clean, structured, almost too good. I read it twice, walked back into the trauma bay, and presented the case to my attending as if I had been thinking about it the whole time.
He asked me one follow-up question about why the superior shoulder suspensory complex (SSSC) mattered for that specific fracture pattern, and I froze. I knew the words. I did not know the reasoning behind them.
That moment stayed with me. Not because I had used artificial intelligence (AI), but because I had used it badly.
I am a third-year orthopaedic resident, and like almost everyone in my program, I find that AI is now part of how I work. We use it to draft discharge summaries, structure case presentations, polish manuscripts, and yes, sometimes to remind us of a classification at 2:00 in the morning. It is fast, available, and confident. Especially the confident part.
And that is exactly where the problem starts.
After the trauma case, I started paying attention to how I was actually using these tools. The honest answer was uncomfortable. I was asking AI to give me answers I had not earned. I was skipping the part of learning that hurts the most, the part where you sit with a question, struggle with it, look it up in three different places, and slowly build something that becomes yours. AI was letting me sound knowledgeable without being knowledgeable, and I could feel the difference whenever an attending pushed back two questions deep.
So, I made a rule for myself, and I am still trying to follow it. I think first, then ask.
Before I open any chatbot, I write down what I think the answer is, even if it is wrong, even if it is messy. For a Schatzker type-IV tibial plateau fracture, I sketch out my mental plan: approach, fixation strategy, what I am worried about. Then I ask AI, and I treat its answer as a second opinion from a very well-read colleague who has never operated on anyone. Sometimes it confirms what I thought. Sometimes it corrects me, and that correction sticks precisely because I had committed to an answer first. Sometimes it is plainly wrong, and noticing that is the most valuable part of the exercise.
I have also stopped letting AI write things I do not understand. Last month, I was working on a paper about rotator cuff repair outcomes, and an AI tool produced a beautifully written paragraph about subscapularis tendon healing biomechanics. It cited two papers. Neither paper existed. They were invented, with plausible authors, plausible journals, plausible years. If I had not gone to look them up, those phantom references would have ended up in a submission with my name on it. That was the day I started treating every AI-generated sentence as guilty until verified.
Using AI as a tool and not as an authority, for me, looks like this in practice: I let it help me organize ideas, not generate them. I let it summarize a long article after I have read the abstract myself. I let it suggest a differential, and then I argue with it. I never let it write a conclusion I cannot defend out loud to my chief.
The deeper question, the one I keep coming back to, is how much is enough? How much AI is too much during training? I do not have a clean answer. What I do know is that the operating room is the place where this question stops being theoretical. When I am at the table and the humeral head is not reducing the way the preoperative plan said it would, when the rotator cuff tissue is worse than the MRI suggested, when the patient’s oxygen level starts dropping, there is no prompt to type. There is only what I have actually internalized.
AI cannot feel the resistance of a screw going into osteoporotic bone. It cannot tell me, from the way a glenoid looks under arthroscopy, that the labrum has deteriorated more than expected. It cannot decide, in the half-second it takes, whether to convert to an open procedure. Those decisions are built from years of repetition, of being wrong in front of someone more experienced, of staying late to read about the case I almost botched. Nothing about that process is efficient, and I am starting to think that the inefficiency is the point.
I am not anti-AI. I will keep using it. It has genuinely made me a faster writer, a more organized presenter, and an occasionally less ignorant trauma resident at 2 a.m. But I have stopped pretending it is neutral. Every time I outsource a thought, I am also outsourcing a small piece of the surgeon I am trying to become.
My residency has not given me a formal class on this. Most have not. So, we are figuring it out in the hallways, between cases, in group chats with other residents who are quietly wondering the same thing. Maybe that is fine. Maybe this is one of those things each generation of trainees must negotiate on its own.
But if I had to leave a single sentence for the resident starting a new year in their program, it would be this: use the tool but make sure that when the OR door closes behind you, the person scrubbed in is still you.
Because in there, AI does not make decisions. We do.
Ruben Arias, MD, is a third-year orthopaedic surgery resident at the Universidad Pontificia Bolivariana in Medellín, Colombia.
Dr. Arias notes that AI tools were used for this post solely to assist with language editing, grammar refinement, and readability. All ideas, opinions, interpretations, and final content were developed and reviewed by the author.
