Med Students Are Choosing the Wrong AI Specialty

If you're a medical student picking a specialty in 2026, you're being asked to bet a decade and $300,000 in debt on a market nobody is teaching you to read.

This episode is the framework I wish someone had given me — plus the data nobody else is putting in front of med students.

In this episode:

• Why 76% of all FDA-cleared medical AI targets a single specialty — and why residency applications to it are still up 30%
• The four kinds of physician cognition (and why "AI replaces cognitive work" is the wrong frame entirely)
• Which specialties get repriced, restructured, or protected — and why psychiatry might be safer than radiology
• The 2026 CMS efficiency adjustment nobody is talking about — and why it's a ratchet, not a one-time cut
• What medical schools are (and aren't) doing to prepare the next generation of physicians for an AI-driven workforce

Watch the video version with charts and visuals: https://youtu.be/NETog9SBtZs

Full sources and the deep dive: danmccoymd.substack.com

Transcript

SPEAKER_00 0:00

Here's a question that should be keeping every medical school dean in America awake at night. A third-year medical student walks into your office and asks, which specialty should I choose if I don't want AI to destroy my career? And you, the person responsible for their education, their $300,000 in debt, their next decade of training have no good answer. Because the honest answer is complicated. And the institutions that train physicians are doing almost nothing to prepare for it. The mainstream narrative goes something like this AI will help doctors not replace them. That's a lovely sentiment. It's also the kind of half-truth that gets people hurt. AI probably won't take your job, but it might take something more important, your pricing power. To understand what's actually happening, you need to look at a single line buried in the 2026 Medicare fee schedule, final rule. CMS finalized a 2.5% reduction to work RBUs. That's the units that determine how much physicians get paid for almost all non-time-based codes. The rationale was broad. Evolving medical practice, better tools, accumulated efficiency gains over time. CMS didn't name AI specifically, but they didn't have to. They built a mechanism, and that mechanism doesn't care whether your efficiency came from AI, from better software, or from a faster internet connection. The result is the same. You get faster, you get paid less. And this wasn't a one-time adjustment. CMS signaled that additional efficiency cuts could come every three years. That's not a policy tweak. That's a ratchet. And AI is about to hand CMS all the evidence it needs to keep turning it. Now I've talked about the RV system before, and I'm not going to relegate that here. That's on Substack if you want it. What I want to focus on today is something different. Not the payment mechanism, but the physicians themselves. Because this efficiency adjustment doesn't hit everyone equally. It creates winners and losers, and the dividing line is more interesting than you'd expect. Here's the part that most people get wrong. They assume the specialties most at risk from AI are the ones where AI can replace the doctor. That's not how this works. The specialties most at risk are the ones where the physician's core work is digitized pattern recognition, work that an algorithm can do faster, cheaper, and at scale. Radiology sits at the top of that list. 76% of all FDA-cleared AI algorithms today target radiology. Not 30%, not half, three-quarters. And this isn't theoretical. A Lancet oncology study of AI-supported mammography screening in Sweden found that AI reduced radiologist reading workload by 44% while maintaining cancer detection rates. Nearly half the work gone. Now the radiologist isn't being replaced tomorrow, but the volume of work that requires a radiologist is shrinking. And here's where the efficiency adjustment bites. VMG Health found that radiology and imaging services already took the steepest overall RV reductions in recent fee schedule updates. They were averaging a decline of about 1.33 total RV use per service. That's a blended number across work practice and expense and malpractice components, not a clean per study work RVU cut, but the direction is clear. Radiology is where CMS is applying the most pressure, and they have the mechanism to keep cutting every three years. Pathology is close behind. The AI pathology market sits around $110 million today, with forecasts ranging from roughly $500 million to over a billion by 2035, depending on who you ask. Whole slide imaging and computational pathology can automate 20 to 40% of screening and quantification work. Not the complex tumor classifications, at least not yet, but the high-volume bread and butter diagnostic work that keeps pathology groups solvent. That's exposed. Dermatology is a more interesting case. About 60 to 75% of dermatologic work is cognitive. Visual diagnosis, pattern matching, AI is already competitive with board-certified dermatologists on image-based lesion classification. But here's the thing about derm. Procedures are a huge part of the revenue. MOSE surgery, biopsies, cosmetic procedures. The AI can look at the lesion, it can't excise it. So derm has a natural floor that radiology and pathology don't. Now, here's where it gets counterintuitive. If you just look at the numbers, psychiatry should be the most vulnerable specialty in medicine. It's 90 to 95% cognitive work, almost no procedures. The entire practice is conversation, assessment, and decision making, exactly the kind of work large language models are designed to do. And yet psychiatry may be one of the most AI-resistant specialties. Why? Because the therapeutic alliance, the relationship between doctor and patient, isn't a nice to have in psychiatry. It is the treatment. A patient disclosing suicide audience to a chatbot and disclosing it to a physician who has treated them for three years are fundamentally different clinical events. The information is the same. The meaning is not. Palliative care, geriatrics, and addiction medicine, they all share the same protection. The clinical work is almost entirely cognitive, but the cognition is deeply relational and embodied. You can't automate a goals of care conversation. You can't outsource the judgment of whether an 87-year-old's I'm fine means she's fine or means she's given up. This reveals an important distinction that the AI will replace cognitive work crowds keeps missing. There's a difference between digitized pattern recognition cognition, the kind radiology does, and contextual relational, embodied cognition. Both are thinking. Only one of them is easily automated. There's a deeper dive on the cognitive work taxonomy and some fascinating data on how different specialties map onto it. That I'll put in the Substack post for this episode, and that link's in the description. If you want job security from AI, learn to use your hands. That's the conventional wisdom. And for once, the conventional wisdom is roughly correct. A curious review of 25 studies from 2024 and 2025 found that AI assisted surgery produced a 25% reduction in operative time, a 30% decrease in interoperative complications, and a 40% improvement in surgical precision. Those are remarkable numbers, and every single one of them describes AI making surgeons better, not making surgeons unnecessary. Neurosurgery, cardiothoracic, vascular orthopedics, these specialties require fine motor skills, control, real-time adaptation to unexpected anatomy, and the kind of situational awareness that no current AI system today can replicate. The robot holds the camera, the surgeon makes the final call. But, and I want to be careful here, protected doesn't mean unchanged. AI is already transforming pre-operative planning, interoperative navigation, and post-operative monitoring. The surgeon's role isn't shrinking, but it is narrowing to the parts that only a human can do. Everything around the procedure, the documentation, the planning, the follow-up protocols, that's all being automated. Emergency medicine and critical care have a similar dynamic. The procedural, high-stake, split-second decision making is resistant, but the documentation, the triage protocols, and care coordination around those decisions, that's fair game. So this is the section that I find most fascinating because primary care is experiencing something that doesn't fit neatly into the replaced versus non-replaced framework. Primary care is being fundamentally restructured. And the data on why is unsettling. Google's AMIE diagnostic AI, tested in a randomized double-blind crossover study published in Nature, outperformed primary care physicians on 30 out of 32 specialist-rated diagnostic axes. In an urgent care feasibility study, its differential diagnosis contained the correct final diagnosis 90% of the time. Eric Topal, one of the most careful voices in medical AI, wrote a piece titled, When Doctors with AI Are Outperformed by AI Alone. Not complemented, outperformed. That's the competitive reality primary care is facing. The traditional PCP model, you feel sick, you make an appointment, you see the doctor for 15 minutes, and you leave, is giving way to something that looks more like a control tower than a clinic. Think about five layers of AI-enabled work that are converging on primary care right now. Layer one, visit level AI. This is ambient scribing, differential diagnosis, support, order suggestions. This is already widespread. Layer two, inbox and admin AI. This is refill triage, normal result messaging, prior authorization prep. This eliminates hours of work daily. Layer three, panel management, AI scanning your entire patient panel for overdue screenings, rising risk scores, care gaps, and generating outreach cues. Layer four, remote monitoring, blood pressure trends, continuous glucose data, weight changes, medication adherence patterns, all synthesized and flagged. And then finally, layer five, social determinants, transportation barriers, missed appointments, food insecurity, identified and routed before they become medical crises. The PCP in this model isn't seeing patients all day. They're orchestrating a system, deciding which AI-generated actions are safe to automate, which need a nurse or a pharmacist, and which actually require the physician to intervene. And this isn't theoretical. Kaiser Permanente logged over 300,000 AI-assisted patient encounters in just 10 weeks, with nearly a thousand physicians using it heavily. Early outcome data from AI-driven care management programs and primary care using algorithm-generated chase list for risk-stratified outreach shows significant reductions in acute events and avoidable hospitalizations compared with match controls. The specific magnitude vary by system and methodology, but the direction is consistent. Proactive AI-enabled panel management produces better outcomes than reactive visit-based care. But here's the uncomfortable part. This orchestrator role is genuinely more valuable than the old model. Population health management prevents disease instead of just treating it. And yet the current payment system has no idea on how to compensate it. An AI agent that identifies 47 missed medication refills, flags 12 pre-diabetic patients for intervention, and catches three dangerous drug interactions generates exactly zero RBUs under the current system. The physician orchestrating all of that gets paid the same as one who didn't even bother. So you'd think, given everything I've just described, that the institutions responsible for training the next generation of physicians would be moving aggressively to prepare them. You would be wrong. As of April 2026, there is no ACGME recognized AI specialty or subspecialty in medicine. None. Zero. And here's a data point that crystallizes the problem. The average number of residency applications per radiology applicant jumped about 30% between 2018 and 2023. Diagnostic radiology positions continue to fill in the mid to high 90% in the match. Students are pouring into a specialty where three-quarters of FDAA cleared AI algorithms are aimed directly at automating the core work. Jeffrey Hinton told the world in 2016 to stop training radiologists. He was wrong on the timeline, but the direction of travel hasn't changed, and nobody in the admissions office is having that conversation. Yale has probably gone the furthest with their AI and Innovation in Medicine distinction pathway, but the program itself has explicitly says it's a distinction, not a subspecial or formal certification. It's a voluntary track within internal medicine residency. Students learn some Python, they do some AI journal clubs, they shadow clinicians using AI tools. It's a good start. It's also wildly inefficient for the scale of what's coming. The ACGME's 2026 annual conference featured a session called Supervising Resident AI Use Without Losing the Learning. That title tells you everything about where the establishment's head is at. They're worried about residents outsourcing their clinical reasoning to Chat GBT, which is a legitimate concern. But it's like a horse breeder in 1910 worrying that students aren't learning proper saddle technique. The entire mode of transportation is changing. Meanwhile, Congress proposed a grant program for AI training at medical schools. The authorization, $1 million per year for five years. That's the federal government's answer to the most significant transformation in medical practice since antibiotics. $1 million a year. The average medical school spends more than that on parking lot maintenance alone. Here's what we actually need: a new residency pathway. Call it clinical AI medicine, I don't care. Call it health systems intelligent, call it whatever you want. That trains physicians not just to use AI tools, but to evaluate them, implement them, manage their failure modes, and redesign clinical workflows around them. We need physicians who understand both the medicine and the technology well enough to be the adult in the room when AI systems actually fail, because it will fail. And when it does, you want a physician in that seat, not an engineer. And there's even a bigger workforce question underneath this one. The AAMC projects the U.S. could face a shortage of up to 86,000 physicians by 2036, 20 to 40,000 in primary care alone. AI isn't a substitute for training more physicians, but it does change which physicians we need. If AI can cut mammography reading workload by 44%, as a Swedish trial demonstrated, and automate a significant share of pathology screening, do we need the same number of radiologists and pathologists we're currently training? Or do we need more AI fluent primary care physicians, more clinical informaticists, more physicians who can govern these systems? The GME dollars exist. We spend roughly $16 billion per year on graduate medical education. The question is whether we're willing to redirect some of that toward the specialties the healthcare system will actually need in 10 years instead of the specialties it needed in 1990. If you're a medical student choosing a specialty right now, here's how I'd think about this. And I want to be honest, I hold this with appropriate uncertainty. First, don't pick a specialty solely based on AI resistance. Pick something you're genuinely passionate about, but go in with your eyes open. If you're drawn to radiology or pathology, understand that your career will look fundamentally different from your attending's. The volume work that currently generates reliable income will be increasingly automated. Your value will shift toward complex interpretation, quality oversight, and AI system validation. That's still valuable work. It's also fewer positions at likely lower per unit reimbursement. If you're drawn to primary care, understand that you're not signing up for 15-minute office visits for the next 30 years. You're signing up to be a population health orchestrator. The good news is that that role might be more intellectually satisfying than the current model. The bad news is that the payment system hasn't caught up yet. And historically, payment reform in medicine moves at geologic speed. If you're drawn to surgery, you're in the best near-term position, but the premium you command will increasingly be for the procedural work itself, not the cognitive work around it. And then there's the liability question nobody has answered. Dr. Sarah Matt put it, if AI is shaping clinical decisions, liability can't sit 100% with the physician holding the mouse. Right now it does. The physician signs off, the physician bears the malpractice risk. But if 30% of your diagnostic workflow is AI generated and you're reviewing 200 AI suggestions per shift at 30 seconds each, is that meaningful oversight or legal fiction? Here's the uncomfortable macro picture. The AMA's own taxonomy now classifies AI-enabled services into three categories: assistive, augmentative, and autonomous. As AI moves from assistive to augmentative to autonomous in each specialty, the physician work component of RVUs, the thing that determines your income, it gets compressed. Not eliminated, compressed. The question for every specialty isn't will AI replace me, it's how much of my current work will AI make visible, measurable, and therefore repriced. The through line here is deceptively simple. AI doesn't eliminate physicians. It makes the cognitive components of their work transparent, and transparent work gets repriced. The specialties that survive best aren't necessarily the ones that resist AI the hardest. They're the ones that find a new source of irreplaceable value, whether that's the surgeon's hands, the psychiatrist's therapeutic relationship, or the primary care physician's ability to orchestrate an increasingly complex system of AI-enabled care. The question medical education needs to answer and answer soon isn't how do we protect residents from AI? It's how do we train physicians who are more valuable because of AI, not less. Because right now, we're minting physicians for a healthcare system that's already disappearing. And at $1 billion a year in federal AI training grants, we're not even pretending to do anything about it. If you found this useful, hit subscribe. The research sources and additional analysis are on my Substack. Links in the description.