We Handed Them a Ferrari and Never Taught Them How to Drive

What 40 future clinicians told us about AI. Co-written with Jhaimy Fernandez, MD, my Alma First co-founder, whose metaphor gave this essay its name. From our February listening sessions.

There’s a moment when you understand that the students are not waiting for us.

Mine came in February, at a dinner table. Fourteen people: a residency program director, physician faculty, a workforce development expert, and five pre-health students. We convened them through Alma First, the nonprofit I co-founded with Dr. Jhaimy Fernandez, to ask a question that sounds simple. What do future clinicians actually want to learn about AI?

One of the students described building her own study system with AI. Flashcards, practice questions, explanations tuned to how she thinks. Nobody taught her this. Her professors tell her not to use it.

Then a physician at the table described being fact-checked mid-visit by a patient holding an AI answer. She said the experience made her more humble. The information is evolving, she told us, and now her patients can look it up too.

Same tool. Same month. Two ends of a career, and nobody in the room had been trained for either moment.

If you’ve been reading for a while, you know this question has been following me. It ran through the Alma First Roundtable with Dr. Karim Hanna on training physicians for the AI era, and through the roundtable with Andrew O’Malley, PhD, on the AI curriculum, whose work sits exactly at the intersection of AI and medical education. Both conversations kept circling the same missing voice: the people still in the pipeline. So this time we asked them directly.

What actually changed

We paired the dinner and a companion webinar with a survey. Forty respondents across the training continuum: pre-health undergraduates, post-baccalaureate and gap year students, medical students, residents, a handful of nurses and career changers. Not a validated national sample. Sensitizing data, gathered on purpose from voices that rarely make it into these studies. We even modeled part of the survey on Offcall’s 2025 Physicians AI Report, so the trainee data could sit next to the physician data and the two could talk to each other.

The largest group, at 40%, was post-bacc and gap/professional year students. Almost nobody surveys them. They are deciding, right now, what kind of clinicians to become.

Here is what changed, structurally.

AI is already inside their training. 77.5% use it at least weekly. 30% use it daily. Exactly one respondent said never.

The education has not followed them there. 30% reported no formal training in statistics, informatics, computer science, or AI. They are learning by improvising, in the gap between professors who say don’t and clinical environments that increasingly say you must.

One student at the webinar named the whiplash exactly. In undergrad, using AI meant you were a cheater. “It’s almost like we were brainwashed a little bit into being like, let’s not even consider opening that app.” Then the same students walk into clinical settings where ambient AI scribes are standard equipment, and the guilt is supposed to evaporate overnight.

This tracks with the national picture. Association of American Medical Colleges (AAMC) data indicate that 86% of undergraduates already use AI for schoolwork, and the 2024 AAMC SCOPE survey found 67% of medical schools have added AI-related content, which means a third still haven’t. Meanwhile the clinical environments really are saying you must. The American Medical Association (AMA) AMA’s 2026 physician survey found 81% of physicians now use AI professionally, double the 2023 rate. The 2025 Offcall Physicians AI Report found 67% of physicians use AI daily, while 81% are dissatisfied with how their organizations are handling adoption.

We spend a lot of energy debating whether AI belongs in medical education. That debate is over. It ended quietly, without a vote, sometime before February.

Where risk shows up first

The number that stays with me is the verification gap.

When AI gives a confident answer, 25% of our respondents always verify it. The rest verify sometimes, rarely, or never.

At the dinner, a pediatrician described her verification habit: run the same clinical question through two different AI tools and compare. The week before, she had done exactly that for a gray-area case. One output read like an emergency. The other said it could wait. Same child, same prompt. Her conclusion has stayed with me: information is cheap now. Knowing whether to trust it is the part that got harder.

These are the people who will be practicing in ten years, and the trainees behind them are forming the habit of checking, or not checking, right now, during flashcards and study summaries, long before a patient is on the other end of the answer.

And the risk is not distributed evenly. During the dinner, one participant put it plainly: students using free AI tools get qualitatively different outputs than students who can pay for premium access. Better explanations. Better study materials. Better preparation. This starts before college. High schoolers studying for the SAT with premium accounts are already getting better answers than the kids on free tiers. And the free tiers, one participant noted, are about to carry ads.

We already knew the pipeline was narrow. In California, only 5% of physicians are Black and 6% are Latino. Now layer a subscription tier on top of it.

Why this creates ethical tension

The trainees in our data are not naive. They are asking the same questions more seasoned clinicians are asking.

Their top concern was losing clinical judgment or professional autonomy. 40% worry AI could replace their thinking if they rely on it too much. They used the language of de-skilling before we introduced it.

They are not alone in that worry. In the AMA’s 2026 survey, 88% of practicing physicians expressed concern about skill loss among trainees. The generations agree on the problem. Neither has been handed a plan.

So the tension is not enthusiasm versus caution. The tension is that we have left judgment formation to chance. When a trainee learns to accept an unverified answer at 11pm before an exam, that is a clinical habit being formed. Nobody consented to that curriculum. It is simply the one running. When the question of consent came up at the dinner, one participant said it would be nice to think we still have the option. Another answered: I think we already passed that.

A nurse of eighteen years told the story that stayed with me. In training, she was made to hand-calculate pediatric IV drip rates, counting drops, and everyone hated it. We have machines for this, they said. Then came the last twelve months. First a wildfire cut her site off from its EHR, down to manual blood pressure cuffs and memory. Then a major health system’s EHR went down and clinics were seeing hundreds of patients a day on paper.

“Don’t tell me, [she said], that you don’t still need to learn how to do this stuff. It worries me that we are offloading some of that”.

Twice in one year. The question under the question: which cognitive skills are we quietly letting atrophy, and who decided?

The equity lens, operationally

Who carries more risk here? The pattern is consistent.

Safety-net populations are treated by clinicians trained with the least AI support, using tools built on data that underrepresents their patients. A physician at the dinner who practices mostly in Spanish described her ambient scribe as good, except for the things it gets wrong in exactly the visits her patients need documented most carefully. So she edits everything. “I’m not writing anymore, but I’m editing.” The typing went down. The vigilance went up.

Underrepresented learners are more likely to be on the free tier, less likely to be at institutions with AI curricula, and more likely to be excluded from the studies that shape national policy. Nearly all published AI education research comes from large, well-resourced academic centers.

Our respondents saw this without prompting. 47.5% named health equity as one of the things they most want to learn about. That did not come from us. It came from them.

How I now think about this

At the webinar, Jhaimy shared a metaphor that has stuck on my mind since then. When we trained, she kept a notebook in her white coat. That is why the coat has so many pockets. If she did not know something, she looked it up between patients. Today’s students pull a phone out of the same pocket and ask the algorithm.

AI feels like being handed a Ferrari, she wrote afterward. The students have it, and they are learning to drive at full speed. Part of us wants to stand behind them and say go faster. But we are also in clinic, and we want them to have the foresight to steer it toward the communities that need them, and the training to know when to brake. Nobody is teaching either skill. That is the gap.

I’ve stopped asking whether trainees are ready for AI. I now ask whether we are ready for trainees who are already ahead of the curriculum.

The most cited learning need in our survey was not prompt engineering. It was how to recognize bias, errors, and safety risks: 72.5%. Then how to use AI without undermining their own clinical judgment: 60%.

Read that again. The students are asking for exactly the education we keep saying they need. The gap is not demand. The gap is supply.

The supply is starting to appear. The team at Offcall and MD+ recently released The Physician’s Guide to AI, a free e-book on AI literacy for physicians and trainees that became a bestseller in its category almost immediately. That speed tells you something: when someone finally builds for this hunger, people show up. Our data suggests what they are hungry for.

The Offcall report tells us what practicing physicians do with AI. The AMA survey tells us how fast that curve is climbing. Our data is the prequel to both. The habits, fears, and verification practices of the next generation are being written now, upstream, mostly unsupervised. If we want different physician data in 2035, this is where it gets decided.

Digital health pearls

A few lines I’ve found myself repeating since February.

The debate about whether trainees should use AI ended before we noticed it started.

Verification is a clinical habit, and habits form during studying, not during practice.

Information is cheap now. Knowing whether to trust it is the expensive part.

The digital divide now has a subscription tier.

One educator at the table predicted oral exams will come back. If AI can write the answer, the assessment moves to the reasoning.

Curriculum by improvisation is still a curriculum. It’s just one nobody designed.

TL;DR

• 77.5% of the future clinicians we surveyed already use AI weekly. Most were never taught how.

• Only 25% always verify confident AI answers. One pediatrician ran the same case through two tools: one said emergency, one said wait.

• Their top request is bias and safety training, not tool tips.

• Free versus paid AI access is becoming a new pipeline inequity, starting in high school.

• Practicing physicians agree: 88% worry about trainee skill loss. The gap is a design problem, and it is fixable.

An invitation to compare notes

This data is small by design. Forty voices, two listening sessions, one evening in February. It generates questions more than answers.

But I keep returning to the student with her self-built study system, and the physician fact-checked by her patient, sitting at the same table, describing the same technology from opposite ends of a career.

If you teach, precept, or supervise anyone earlier in training than you, try one question this week: what are they doing with AI that you haven’t asked about yet?

I suspect the answer is more than you think. It was for us.

And if this essay named something you’ve been feeling, send it to someone earlier in training than you. They are the ones this conversation keeps leaving out.

This essay draws on listening sessions and survey data gathered by Alma First, the 501(c)(3) I co-founded with Dr. Jhaimy Fernandez, whose vision shaped this work from the beginning. Participant remarks are shared by role only, consistent with the ground rules of those sessions. Full findings will be available through Alma First upon request.

Disclaimers: All views expressed are my own and do not represent my employer or any institution I am affiliated with. Any tools, products, or technologies mentioned are included for educational purposes only and are not sponsored or endorsed. Nothing in this piece should be interpreted as medical advice.