Why I Did Apply to Residency

The first time I worked nights as a resident, I held the pager for every surgical service in the hospital, consults included. Around 40 patients. In all of medical school, I never held a pager that mattered.

A few months before that night, Aditya Jain's Doximity essay opens with a scene I keep coming back to. An AI platform has read the chart, taken the history, built the differential, and drafted the note. The physicians' only job is to press "approve." The patient gets their meds and leaves happy.

He's right about the button. I think he's wrong about what it takes to press it.

I spent medical school building AI for medicine: Dendro, which turns lectures into study tools, and research testing LLMs on clinical work. So I don't need convincing that AI is coming for the cognitive core of medicine. I still think residency was the right call, and the AI is a big part of why.

Residency trains two things that button depends on: the judgment to know when the AI is wrong, and the firsthand knowledge to build AI worth approving.

Three months isn't much. Someone will tell me to check back at PGY-3. That's the point: if three months showed me this much, imagine skipping all of it.

What he gets right

A lot of his case holds up. Residency is heavy on service; admitting the fifth "weak and dizzy" patient at 4 a.m. is not deliberate practice. Academia is slow. The benchmark curves are real: on Radiology's Last Exam, a generalist model went from 30% to 51% in about a year, passing radiology residents.

Where I part ways is the frame. He treats residency like a bond: you pay in years, you get expertise back, and if AI makes the expertise worth less, the bond is junk. That only works if residency's product is the differential diagnosis. It isn't. An oncologist, Roupen Odabashian, puts diagnosis and decision-making at "maybe ten percent of my day" (Odabashian). The benchmarks measure that ten percent. Residency trains the rest.

1. Residency hands you medicine

As a student, you present and someone else decides. Your orders get cosigned. The 2 a.m. page goes to someone else.

In residency the page comes to you. On my first day I held the trauma floor pager. Since then I've seen patients in the ED and gotten them to the OR, run to a rapid response where what we did in the first few minutes stabilized the patient, and sat with patients who just needed someone to listen.

The difference isn't volume. Every decision now has your name attached to what happens next, and that changes what you notice. Every decision has gravity. No benchmark puts your name on the outcome.

It also changes what you'd build. Once you've owned the decision, you know what a tool has to get right at 2 a.m. with forty patients on your list, and what it can never get wrong. Someone who has only seen the workflow in a demo designs for the demo.

2. Someone has to know when the AI is wrong

In Jain's future, the physician presses approve. The safety of that whole system lives in the rare moment they don't.

A group of clinician-educators writing in BMJ Digital Health put the problem in one sentence: a generated conclusion "can only be verified by someone who could have reached it independently" (BMJ Digital Health). The approve button only works if the person pressing it could have written the plan themselves.

That's why skipping training doesn't escape the AI problem. It walks straight into it. Researchers now name three harms: deskilling (losing skills you had), mis-skilling (learning the AI's errors as truth), and never-skilling, where AI does the work before you ever learn it.

Never-skilling is the one that should worry my generation most. Deskilling happens to doctors who already have the skill. Never-skilling happens to everyone after them. AI is available from the first day of medical school, before any clinical reasoning has formed (Nature Medicine). A student who asks for the differential before building their own gets an answer that feels like knowing it. Repeat that across a clerkship, then a career, and you get someone who can approve a plan but couldn't have made one. "Just verify the output" doesn't fix it, because verifying takes a standard to check against. Someone who couldn't have produced the answer can only accept it (BMJ Digital Health).

A doctor who never trained isn't deskilled. They were never skilled. As a Nature Medicine Perspective put it this year, "a co-pilot is only useful if trainees first learn how to be pilots." Skipping residency is never-skilling taken to its limit: the training doesn't get thinner, it never happens.

Deskilling is already measurable. After routine AI-assisted colonoscopy, experienced endoscopists' adenoma detection without AI fell from 28.4% to 22.4% (Lancet Gastroenterol Hepatol00133-5/abstract)). But look at who gets fooled. When an AI suggested the wrong category on mammograms, inexperienced radiologists got those cases right less than 20% of the time. Very experienced radiologists got them right 45.5% of the time (Radiology). Experience didn't make anyone immune. It made them much harder to fool.

The human isn't the weak link. The untrained human is.

You build that standard on real patients, while someone checks your work. A consult came in to rule out appendicitis. The question arrived already framed, the way an AI's answer does: here's the likely diagnosis, confirm it. Once you looked at the whole picture, it was clear the cause was a malignancy. Answering the question as asked would have been easy, and wrong. The job was to notice that the question itself was wrong.

That is the exact skill the approve button depends on. A model handed "rule out appendicitis" will usually answer the question it was asked. Learning to question the question takes reps on real patients, while someone is still checking.

3. What AI can't be in the room for

Jain predicts the doctor becomes "a customer service representative with liability," the person who signs. That misses what patients actually come for. Odabashian's patients already ask ChatGPT before they see him. What they come for, he says, is to be helped to decide "by someone who will still be there when the decision goes badly."

One of my patients was sixteen days out from an exploratory laparotomy for a small bowel obstruction. His bowel function still hadn't come back. He was sad, and starting to sound defeated. There was no order I could write to fix that. So I sat down. We talked for a while. I learned about his life, and I told him how far he'd already come.

Jain would agree that matters. If patient care gives your life meaning, he says, do residency. And "if you can stomach it, we do need physicians to train AI."

But that frames the conversation as a personal preference, something for the people who happen to enjoy bedside work. It isn't a preference. It's part of the care, and it's a part AI isn't in the room for. What that patient needed wasn't information. A model could have told him that ileus after a big abdominal operation usually resolves. He needed a person sitting next to him who would still be there tomorrow. Odabashian makes the same point about the research: the studies pitting AI against doctors measure his ten percent, and "the other ninety percent is not in the literature at all."

If you want to build the systems that will sit between patients and doctors, you have to understand the part of medicine those systems will never be present for. The only way to learn it is to be the one in the chair.

4. You find the problems by being in the room

Jain describes residents muttering that someone should fix a broken workflow, then moving on because seven more patients are waiting. He reads that as residency killing curiosity. I read it as the only place those problems are visible.

From outside, most workflows look fine. The protocol exists, the order set exists, the dashboard is green. The problems live in the gap between what the system says should happen and what actually happens, and you only see that gap when it lands on you: the patient still NPO because the OR time moved again, the discharge stuck on one missing piece. On my nights, maybe two in five pages truly needed a doctor. That ratio isn't in any dataset. I know it because I was holding the pager.

That's the builder's case for residency. The person muttering on rounds isn't the problem. They're the only one who has seen it clearly. The best healthcare AI won't come from people who studied the problem. It'll come from people who had it first.

5. The people in the building

Jain is right that academia moves slowly. But the decisions that matter for AI in medicine aren't being made in journals. Hospitals are deciding right now which tools to trust, and in Doximity's own survey, only 8% of physicians said their institution's AI policies are clear (Doximity). Those policies will be written by attendings, CMIOs, program directors, and nurse managers: the people residents work next to at 2 a.m.

In three months, residency has put me across the table from the people who built an AI tool now used across our entire health system, from the founder of one of the largest surgical education platforms, and from the AMA, which interviewed me about our study on AI and studying. None of that came from a cold email. It came from being in the building.

For a builder, that's not networking. It's the trust and the access you can't buy later: the clinicians who will tell you the truth about your tool, and the institutions that will actually let you deploy it.

6. Even if he's right about the curve

Suppose he's right. Suppose by 2030 the model beats me at diagnosis.

Then the diagnosis stops being scarce. What becomes scarce is everything around it: the person who can tell when the model is wrong, who knows what this patient can tolerate, and whom the patient trusts with the decision. "When information becomes free, judgment becomes valuable" (source).

That breaks the bond math. Jain prices residency as an asset that loses value as AI improves. I think it runs the other way. The better the AI gets, the more the check is worth.

We've run this experiment once. In 2016, Geoffrey Hinton said to stop training radiologists. There are now more than 700 FDA-cleared radiology AI models (Understanding AI). Radiology residency positions still rose 33% from 2010 to 2025, and radiologists per capita haven't budged, because demand outran supply (JACR). Jain's purest cognitive specialty is the one that proved him wrong first.

The "physician and" trap is real. Residency isn't what springs it.

Jain's best structural point: by the time you reach the "and" in "physician and," you're 35, in debt, and wired to avoid risk. I take that seriously. Some weeks the building gets almost none of me.

But the trap isn't residency. It's stopping building during it. Three months in, a paper I co-authored on AI-generated study material came out in BMC Medical Education (paper). Dendro is still running. This essay exists.

When Josh Farkas called medical school "trade school for physicians," Jain replied that residency is the actual trade school (thread). I agree with him. That's why I'm here.

AI will do the differential. That part isn't really in question. What's in question is who will be able to tell when it's wrong, and who will build the tools worth trusting.

Skip residency and you can build the AI. Do residency and you can tell whether it works.

If you skipped residency to build, or you're building through it, I want to hear what it taught you that a benchmark couldn't.