Health singularity, a clinic hallway, and a question I keep hearing
Last Tuesday, after a long clinic session, a resident stopped me in the hallway and asked the question that has become almost routine: “If AI keeps getting better, why would anyone still go to medical school?” Earlier that afternoon I had watched a patient with diabetes hand me a phone full of glucose logs, medication photos, and a note from her daughter about what changed after the last visit. The data were abundant. The clinical judgment still had to be earned. I left that day thinking about how much of medicine is invisible to people who only see the facts, and how much of it is really about pattern recognition, prioritization, and staying calm when the answer is not obvious.
The health singularity does not make medical school obsolete. It makes medical education more valuable, because the hard part of medicine has never been memorizing facts alone, it has been learning how to reason under pressure, integrate uncertainty, and make accountable decisions with other humans and now with AI systems.
That is why future doctors will need two forms of mastery: the classic clinical habits medicine already demands, and the ability to supervise AI tools with the same discipline aviation uses for automation.
I used to think the rise of generative AI would mainly devalue medical training by compressing the factual work that students spend years learning. Then I watched how quickly AI could produce plausible answers that were shallow, overconfident, or wrong in exactly the places where a novice is least able to catch the error. Now I think the deeper value of medical school is not the encyclopedia, it is the operating system.
Medicine still trains things AI does not own
In my experience as a physician-executive, medicine trains scientific reasoning under constraint. A student learns to build a differential diagnosis from partial data, to notice what does not fit, to integrate labs, imaging, physical exam, trajectory, and context. That is a habit of mind, and it is built the hard way. You repeat it until it is reflexive.
It also trains stamina. Long call nights, compressed decision cycles, repeated emotional exposure, and the requirement to stay useful when tired are not glamorous, but they matter. I think of medical training as closer to an elite academy than a credentialing factory. The Navy SEAL comparison is not perfect, but the underlying point lands: medicine selects for people who can endure, adapt, and perform under stress.
That matters more, not less, in an AI-saturated clinic. If the machine can draft the note, summarize the chart, and suggest the next test, the human’s advantage shifts toward judgment, supervision, and responsibility. A doctor who cannot reason independently becomes dangerous when the model is wrong in a subtle way.
The new baseline is AI literacy, not AI dependence
The aviation analogy is useful here. In Flight rules for clinical AI: lessons from aviation for human-AI collaboration in medicine, Ong and colleagues argue for a model of clinical AI that preserves manual competence, uses simulation for failure modes, and treats overreliance as a safety problem rather than a convenience. They describe the need for minimum unaided practice, concordance monitoring, and structured debriefs. That framework fits medicine because it recognizes a brutal truth: automation can make systems safer and users weaker at the same time.
I see the same risk in education. AI-induced never-skilling in medical education makes the warning explicit, even if it is still a conceptual paper rather than a measured clinical trial. The concern is simple. If trainees let AI carry the cognitive load too early, they may never build the foundation they need to judge the machine later. That is not a minor pedagogic issue. It is a future workforce issue.
There is a useful number from the education literature here. A 2026 Cureus scoping review of conversational AI in medical education identified 20 studies after screening 496 records and found that several reported statistically significant gains in examination performance, history-taking, and structured clinical assessment scores. I do not read that as proof that AI should replace teachers. I read it as evidence that AI can be a strong tutor when the learner already has structure, supervision, and feedback.
Another 2026 survey in BMJ Health Care Informatics found that Swedish physicians and the general public expected medical AI to outperform current human practice, with acceptable sensitivity thresholds reaching 95% for physicians and 100% for the public in a chest pain triage vignette. That is a useful reminder that trust will not come from hype. It will come from performance, and from the visibility of failure modes.
What I would not do
I would not let a trainee use AI as a substitute for learning differential diagnosis, note-writing discipline, or the mechanics of clinical reasoning. I would also not deploy a clinical model into a workflow without clear ownership, escalation paths, and an answer to the question, “What happens when it is wrong at 2 a.m.?” If no one can answer that, the tool is not ready.
This is where the physician-executive lens matters. Hospital AI governance cannot be a vague committee report. It has to address regulatory pathways, validation, change management, monitoring, and liability. In the United States, that means knowing whether a system arrived through FDA 510(k), De Novo, or PMA, and whether the local use case matches the claimed indication. It also means understanding that a polished demo is not a safety case.
I have sat through vendor presentations that made me uneasy for exactly this reason. The demo looked clean. The workflow was “streamlined.” The gap was always between the slide and the shift. The slide never had a pager going off.
Why medicine stays valuable even if the facts become cheap
The most persuasive argument for medical school is not nostalgia. It is that the practice of medicine forces repeated integration across domains that AI can assist but not morally own. The physician learns how to weigh tradeoffs between sensitivity and specificity, when to observe and when to act, when to order the test and when to stop ordering tests. That discipline reduces harm.
Medicine also teaches a form of judgment that is hard to automate because it is social as much as technical. Patients do not bring isolated variables. They bring fear, family pressure, money problems, language barriers, grief, and memory that is incomplete. A model can summarize those variables. It cannot carry the responsibility of deciding what matters most to the person in front of you.
I once told a medical student who was exhausted and second-guessing herself, “You are not just learning facts. You are learning how to stay functional while your brain is on fire.” It sounded blunt in the moment, but it was true. Medical school teaches that kind of resilience, and the profession still needs it.
The future doctor will need something else too. The future doctor will need to be fluent in AI failure, not merely AI output. They will need to know when a model is making a confident mistake, when a workflow is nudging them toward automation bias, and when the safest move is to slow down and think. That is a clinical skill.
There is also a broader societal reason this path still matters. In a world where AI can narrow the futures people consider, as shown in Faith in AI can narrow the futures individuals consider, medicine remains one of the few professions that still demands real-world accountability, close human contact, and repeated exposure to the consequences of getting things wrong. That is a hard school. It should be.
What the health singularity changes
If the health singularity arrives in the practical sense, meaning AI is embedded across triage, documentation, diagnosis support, and operational work, then the value of medical education shifts rather than disappears. Students will spend less time proving they can recite what a model can retrieve instantly, and more time proving they can supervise systems, challenge outputs, and recognize when a human conversation is the real intervention.
That changes curriculum priorities. I would want more training in probabilistic thinking, more simulation with bad AI outputs, more governance literacy, and more time spent on cases where the right answer is not the loudest answer. I would want students to practice saying, “I do not trust this result yet,” and then showing why.
Some people hear that and think medicine sounds less essential. I hear the opposite. The more AI expands, the more valuable it becomes to have clinicians who can separate competence from confidence, and output from understanding.
I do not want a generation of doctors who merely press approve. I want clinicians who can reason, supervise, and refuse when necessary. That is the job.
If you want the credentials behind this perspective, my professional background is summarized on Dr. Sina Bari’s physician profile and training background. I mention that deliberately, because in this debate the pedigree matters less than the practice pattern: have you actually worked where decisions are made, under time pressure, with real consequences?
Back to the hallway
At the end of that clinic day, I ran into the resident again. I told him the answer was not to stop sending smart people to medical school. The answer was to make sure they learn medicine deeply enough that AI becomes an instrument, not an alibi. He nodded the way trainees do when something lands a little harder than they expected.
That is where I have landed too. I used to think the health singularity might shrink the value of medical education. Now I think it will expose which parts of medical education were merely informational and which parts were actually formative. The latter still matter. A lot.
FAQ
Will AI make medical school unnecessary for future doctors?
No. AI will reduce the value of memorizing routine facts by hand, but it will increase the value of clinical reasoning, supervision, and judgment. The doctors who thrive will be the ones who can both think independently and evaluate AI output critically.
What is the biggest risk of AI in medical education?
The biggest risk is never-skilling, meaning trainees may never build foundational reasoning if AI does too much too early. That can leave them unable to detect when the tool is wrong, especially in subtle or high-stakes cases.
How should hospitals govern clinical AI tools?
Hospitals should treat AI as a governed clinical system, not a software novelty. That means validating the intended use, understanding the regulatory pathway, assigning ownership, monitoring performance after launch, and planning for failure at the bedside and after hours.
What is Dr. Sina Bari’s approach to AI in medicine?
My approach is pragmatic and skeptical. I want tools that improve patient care, but I want them supervised, measurable, and integrated into workflows where a clinician still understands the decision, the failure mode, and the escalation plan.
What should a medical student learn now if AI is changing everything?
Learn reasoning, not just retrieval. Learn how to build a differential, communicate uncertainty, check your own thinking, and understand the limits of AI. Those skills will age better than any single tool.