Medicine in the AI Era

Part of Before You Commit. I work in two worlds, one as a hospitalist and one building with this technology. Neither the people promising AI will replace doctors nor the people insisting nothing will change are describing what I actually see.

I work in two worlds. I'm a hospitalist, so I take care of patients on the wards. And I also work in AI: I've helped train clinical models and spent real time benchmarking how well they do the things doctors do. So when a premed asks me some version of "is it even worth becoming a doctor if AI is going to take over," I don't answer from the sidelines. I've seen what these models can do, and I've seen where they fall apart. That's the honest place I'm writing from. I don't have the whole future figured out, and anyone who tells you they do is selling something.

Let me walk you through how I actually think about this, in the order it matters.

The fear

The fear is real and I'm not going to pretend it away: AI will replace doctors, so why sink ten years and a mountain of debt into this?

Here's what I can tell you from where I sit. The data doesn't support that, at least not yet. There is still a deep, growing need for physicians. The AAMC projects a shortage of up to 86,000 doctors by 2036, and that's with AI already in the building.1 The problem this country has is too few doctors.

My sharpest point is one people don't talk about enough. AI has no liability. When a diagnosis is wrong, when a decision goes bad, when a family wants an answer, somebody has to hold that responsibility, sign their name, and stand behind it. Doctors are liability sponges. That's part of the job almost nobody applies for on purpose, and it's exactly the part a model can't absorb. A machine can suggest. A licensed human being has to be accountable. That alone means doctors aren't going anywhere soon.

So I'd call medicine a relatively safe bet. Not a risk-free one. Nothing is. But safe enough that I wouldn't let the AI panic be the thing that talks you out of it. There are plenty of good reasons to think hard about this path, the debt, the years, whether the life actually fits you. "A robot is going to take the job" is not the one I'd lose sleep over.

The real divide is between doctors who safely and efficiently use clinically-validated AI and doctors who don't. That second group, the ones who refuse to touch any of it, are probably going to end up slower and less helpful than the ones who learn to use good tools well. What changes is the bar for what a good doctor does.

And adoption is already enormous. Reported usage of OpenEvidence, one clinical AI tool, runs from about half of U.S. physicians upward depending on who is counting and when, with the company reporting hundreds of thousands of monthly users.2 Treat the precise share as soft, because these are company-reported numbers relayed by trade press rather than independently audited ones. The direction is not soft at all. Your future attendings are already using this stuff between patients. AI isn't coming to medicine. It's here, and it's woven into the workday.

What AI is and isn't good at right now (July 2026)

This is the part where my benchmarking work makes me want to be precise, because both the hype and the doom get it wrong.3

Frontier models today are genuinely, surprisingly good at a few things. Clinical reasoning is one. Differential diagnosis is another. Give a strong model a clinical vignette and it will often build a differential that would make a good resident nod. There's a growing body of serious work showing this, including benchmarking in NEJM AI and diagnostic-accuracy studies published in Nature. Multimodal interpretation has also gotten strong. These models can look at an image and answer clinical questions about it in ways that were science fiction a few years ago. I'm not going to downplay that, because it's true and it's important.

And there's a lot they're not there on yet. ECG interpretation is still imperfect. Some imaging modalities trip them up. Complex, multi-step diagnostic reasoning, the kind where you're holding six competing possibilities and updating them as three new results come back over two days, is still shaky. Natural multi-turn conversation, the back-and-forth of an actual patient encounter where the real story comes out on the third or fourth question, is not solved.

But I'd be lying to you if I said the progress hasn't been real, or that these gaps are permanent. Lab interpretation is improving. The modalities that stump models today are exactly the ones people are working hardest on. If I told you in confidence what these systems couldn't do two years ago, half of that list is gone now. Plan for a world where they keep getting better, because they will.

Doing well on a clean vignette is not the same as doing well in a messy clinic. In a test question, the relevant facts are handed to you in a tidy paragraph. In real life, the patient minimizes the pain, forgets the medication they're on, gets the timeline wrong, and mentions the thing that actually matters as you're reaching for the door. Gathering that story, deciding what to trust, knowing when the numbers don't match the person in front of you, that's most of the job, and it's the part the benchmarks barely touch. So when you see a headline that a model "passed the boards" or "beat the doctors," read it carefully. It's usually true and usually narrower than it sounds.

De-skilling is the risk that deserves your attention

For you, a trainee, the real danger is that you never build the skills in the first place, or that you let them rot.

I'm part of the last generation trained mostly before these tools were everywhere. I built the foundation the slow, painful way, and then the tools showed up. That order matters more than it sounds. I know what a differential feels like when I have to grind it out myself, so when a model hands me one, I can tell whether it's right, and I can tell when it's confidently wrong.

Now picture someone who came up the other way, leaning on AI for the reasoning and the logic from day one. Two things can go wrong. They might never develop the foundational skills at all, because the tool did the hard part every time. Call that never-skilling. Or they build the skills, then lean on AI so hard that the skills fade. Call that de-skilling. And this isn't just my worry. There's real evidence now. A 2025 multicentre study in The Lancet Gastroenterology & Hepatology looked at experienced endoscopists and found that after continuous exposure to AI assistance, their adenoma detection rate on standard, non-AI colonoscopies dropped from 28.4% to 22.4%, a relative decline of about a fifth (⟳ verified July 2026). These were skilled doctors. The tool made them worse at the thing they already knew how to do, once they lost the reps.

Skills you don't build during training, or lose during training, you may never get back. The foundation gets poured in those years or it doesn't get poured at all. And the cost of borrowing your thinking doesn't show up for years, until you're the one holding the responsibility and you reach for a skill that was never really yours. So every premed and med student has to look in the mirror and ask an honest question: how am I going to keep sharpening the real skills while I have every reason to outsource them?

The line I draw is simple. The AI-reliant version of you can be replaced. The version of you who builds real skills and uses AI to sharpen them is the clinician of the future. Same person, two very different futures, and the fork is how you train right now.

And I want to be fair to you, because this is not entirely on your shoulders. The whole system is going to make the easy path very easy. The tools will be right there, free or nearly free, faster than doing it yourself, and no one is going to stop you from leaning on them. That's exactly why you have to be the one who decides where your own line sits, and why you have to protect the parts of your training where you're supposed to struggle. The struggle is the point. Nobody is coming to protect your skills for you.

How to actually use AI as a premed or med student

None of that means don't use it. Use it. I mean that.

AI is a phenomenal study partner. Feed it your notes and have it quiz you. Ask it to find the holes in your understanding, the concepts you keep getting wrong. Have it build you a focused study plan for the three weeks before an exam instead of the scattershot one you'd make at midnight. Reorganize a messy unit into something that makes sense. Turn a dense chapter into a summary you can actually review, or a script you can talk through out loud. If it helps you learn faster and understand deeper, that's a good use of a good tool. The world really is open to you here in a way it wasn't for me, and there's no shame in using it to get better.

The line is one thing, and it's clean. Don't outsource your learning, your thinking, and your training. Use AI to study; don't use it to avoid studying. Use it to check your reasoning; don't use it instead of reasoning. The moment it's doing the thinking so you don't have to, you've stopped becoming a doctor and started becoming a person who can look things up. Use it to build the muscle, not to skip the workout.

Don't let it cost you your humanity

We're humans first. The thing that makes someone want to be a healer in the first place is human, not technical. Don't let the spread of AI quietly cost you that. As more of medicine gets automated, real empathy gets rarer, and rare things get more valuable. Not the flat "I'm sorry to hear that" a chatbot gives you, but the kind that happens in the room, from one person to another.

Think about what AI will never teach you. Grieving alongside a patient. Sitting with someone the hour after they've lost the person they love most. Talking a family through code status when the moment is bad and there's no clean answer. Breaking news that's going to change someone's life before they've finished their coffee. Trying your absolute hardest and watching a patient recover, and trying just as hard and losing them anyway. You don't learn those from a model or a module. You learn them by living them. You have to feel it, mess it up, sit in the discomfort, learn, lose, and win before you turn into an actual physician and an actual healer.

So use AI. Let it keep getting better. Don't let it replace you, and don't let it dull you. Those are two different warnings and you need both.

Should it change your specialty?

Short version: no field is obsolete right now, so don't pick or avoid one out of fear.

Radiology is the case everyone points at, so let's use it. People have been predicting the end of radiology for over a decade, and instead radiologists are in serious demand, salaries have climbed, and the job market is hot. AI in radiology has landed as an adjunct, a tool that helps read faster and handle more volume, not a replacement. If you want the full picture, we lay it out on the diagnostic radiology page. Not one physician job has been fully automated away yet.

I'd rather you hear this from me: that could change. I don't know the timeline and neither does anyone else. And there's a genuinely hard question I want to put in front of you without pretending I can answer it. In places with no doctor at all, the rural clinic four hours from a specialist, the community that has never had reliable physician access, is an AI doctor better than no doctor? Maybe. Nobody actually knows yet, and the legal ground under that question is completely undefined. I'm not going to resolve it for you. I just want you thinking about it, because your generation is the one that's going to have to.

Don't plant your flag at either extreme. "AI will never change medicine" is naive. "AI will replace all doctors" is fantasy. The truth is somewhere in the messy middle, and the people who do well will be the ones who keep watching the tools change, adapt their workflow to match, and never lose the reason they wanted to do this in the first place.

Bottom line

AI doesn't have to be taboo in medicine. I'm not anti-AI, not even a little, I build this stuff. But safe, ethical, clinically-validated use has to come first, every time. The goal is simple: use AI to become a better doctor. Build your own skills. Use AI to sharpen them, not to replace them. Do that, and you're the one the whole thing still needs.

Last reviewed: 2026-08-03. This page is educational, not medical or career advice, and it reflects one physician's honest read of a fast-moving field.

This is the fastest-decaying page on the site, and it is dated on purpose. Capability claims about AI systems, adoption figures, and workforce projections all move faster than an annual review cycle. The capability assessment on this page describes July 2026. Read anything specific here as a snapshot rather than a standing fact, and check the sources before relying on a number.

References

Footnotes

  1. AAMC, The Complexities of Physician Supply and Demand: Projections From 2021 to 2036, released March 2024. Projects a shortage of up to 86,000 physicians by 2036 across primary and specialty care. Earlier editions of this projection gave materially different figures, so cite the edition rather than the number. https://www.aamc.org/news/press-releases/new-aamc-report-shows-continuing-projected-physician-shortage ↩

  2. Adoption figures for OpenEvidence are company-reported and relayed by trade press rather than independently audited, and the reported share of U.S. physicians using it has been stated variously at around half and higher within the same year. An earlier version of this page said "roughly two-thirds" while linking a source whose own headline said half; the text now reflects the range and its softness. Treat the trend as well established and any single percentage as provisional. https://www.beckershospitalreview.com/healthcare-information-technology/ai/openevidence-6-things-to-know-about-the-ai-tool-used-by-half-of-physicians/ ↩

  3. Capability claims in the section above describe frontier models as of July 2026 and are the author's assessment from clinical and technical practice rather than a citation to a benchmark. Published medical-benchmark results move quickly and benchmark performance is a poor proxy for clinical performance, which is the reason this section is dated rather than sourced to a leaderboard. ↩