About This Series
Alma First is a nonprofit co-founded by Dr. Gigi Magan and Dr. Jhaimy Fernandez. We build leadership pathways for communities historically excluded from healthcare and technology. As medicine becomes more digital, we train the next generation of digital health leaders and work to make sure new tools do not widen the disparities they are supposed to narrow. The Alma First Roundtables are a monthly series where physicians and future physicians with different lenses sit with the same questions and answer from where they stand. No consensus required. A real conversation about what matters.
Background
One in five medical students is using AI to write clinical notes. Nine in ten institutions have no formal policy on how, or whether, they should.
The numbers make the gap concrete. A 2024 survey across 192 medical schools found over 75% of students received no formal AI education. A 2025 survey of US osteopathic schools found 93% of institutions lacked formal policy guidelines for student use of generative AI. Seventy-nine percent had no plans to develop them.
Meanwhile, 43% of medical residents now report using generative AI tools daily. Over 66% use them weekly. Thirty-four percent use them for clinical documentation.
The gap is not theoretical. It is happening now, in every medical school and residency program in the country.
This roundtable started with two essays. My co-founder Jhaimy Fernandez wrote about handing students a Ferrari with no driving lesson: prohibition creates secret drivers, not safe ones, and faculty need to teach the turns. I wrote a companion piece about where the risk concentrates when training lags behind the tools. In the visits where AI is most likely to miss something. With the patients who have the least margin for error.
I kept returning to the same question. Not “should we teach AI in medical school?” That ship sailed while we were writing the curriculum proposal. The question is: are we preparing physicians to hold the responsibility these tools carry, or are we leaving them to figure it out alone?
I did not want to keep asking that alone. I wanted to sit with it in a room with people who see the gap from different angles. A clinical informaticist training residents in real time. A future physician about to walk into it.
So I called Dr. Karim Hanna. And I asked Lorena Gonzalez, an Alma First fellow and incoming medical student, to join us.
The Roundtable
Dr. Gigi Magan: Family physician at an FQHC in Northern California. I teach at UC Riverside School of Medicine, research how AI performs in primary care, and co-founded Alma First with Jhaimy Fernandez. I am the person in the room who asks the question nobody wants to sit with: does this tool hold up on the hardest day, in the visit with the fewest resources and the most at stake? If the answer is “we are not sure,” we need to be honest about that before we deploy it.
Dr. Karim Hanna, MD: Family physician, program director, and clinical informaticist at USF Health in Tampa. I write the AI+MedEd newsletter because the educators shaping tomorrow’s doctors need a trusted guide through this moment. My residents are navigating a landscape that didn’t exist when I trained, and helping them do that thoughtfully is one of the most important things I can do.
Lorena Gonzalez: Incoming medical student and Alma First fellow for the past two years. What draws me to medicine is how versatile it is: physician, advocate, leader, educator, all in one career. That platform is a privilege, especially for people whose voices often go unheard. What excites me about AI in healthcare is its potential to expand access for communities historically excluded from it. Multilingual tools helping patients navigate portals in Spanish. Screening tools flagging early disease in places without specialists. What worries me is what happens when we are not careful: bias, privacy risks, and systems that deepen the inequities they claim to solve.
The Questions
Question 1: What is medical education still not teaching well enough about AI, and why does that matter?
Gigi
We are not teaching students where AI is most likely to fail them and their patients.
Most AI education, when it exists, happens in controlled settings. A lecture hall. A simulation center. A well-resourced academic medical center where the tools work as advertised. The gap between what AI does in a demo and what it does in a visit with a Spanish-speaking patient, an interpreter on video, and three unaddressed chronic conditions is enormous. Nobody is naming that gap.
Here is what the research shows.
In one study, AI-generated clinical notes averaged over 23 errors per case. The most common error type was omission, accounting for 86% of all errors.
That is the hardest kind to catch because it requires the clinician to remember what was said, not to read what was written. The more complex the encounter, the worse the tool performed. Accuracy was inversely correlated with transcript length and complexity.
It gets worse. Most ambient AI scribes currently support only conversations conducted entirely in English. A 2025 study found significant disparities in automatic speech recognition accuracy when transcribing speech from African American patients compared to White patients. The same research documented speaker attribution errors: the tool fails to distinguish between multiple speakers. That is what happens in encounters involving interpreters or family members. In the visit I described, the one with the interpreter on video, the tool was not designed for that encounter at all.
In some of the most complex visits in medicine, these tools were never built to work. We are not teaching students to evaluate AI output the way we teach them to evaluate a physical exam finding: as a practiced clinical skill, under real pressure, with the understanding that getting it wrong has a cost. When a student signs an auto-generated note without reading every line, we have not taught efficiency. We have taught uncritical acceptance.
Karim
We are teaching residents what AI does. We are not teaching them what it gets wrong, or why. There is a real difference between a resident who uses an ambient scribe because it saves time and one who knows to re-ask a question when something in the AI-generated history does not make clinical sense. That second learner has something the first one does not: epistemic humility about the tool.
Generative AI is a fast, confident junior colleague who has never seen a patient.
Our job is to supervise it. When we skip those conversations, we risk producing physicians who outsource their clinical reasoning incrementally. By the time the stakes are high, they do not realize how much judgment they have quietly handed off.
Lorena
Still, I do not believe this responsibility should rest entirely on educators, who are navigating this shift in real time as well.
There is a generational gap. Students live with technology daily. Many faculty did not train with it. Updating a curriculum takes work nobody is compensating. Will professors receive additional pay for the extra time required to research and teach a topic with no existing template? Nobody has answered that question.
There is also negativity bias. Students notice both the promise and the risk of AI, but end up fixating on the risk. That matters because AI is going to shape how we practice medicine, and we are not going to figure it out in isolation. Students and educators already carry exams, clinical responsibilities, and competing priorities. If we co-create solutions and share the burden, everyone benefits. If we do not, we risk over-relying on tools we do not understand, or arriving unprepared when patient care depends on sound judgment. As future clinicians, we owe our patients the effort to understand AI regardless of whether we personally like it or feel comfortable with it.
Question 2: What skills, mindsets, or habits do future physicians need to use AI well without losing clinical judgment, humanity, or trust?
Gigi
The skill nobody is naming: the ability to slow down when a tool speeds you up.
AI makes certain tasks faster. Documentation. Literature search. Differential generation. The efficiency is real. But speed introduces a new risk. When the output arrives pre-packaged, the temptation to accept it without interrogation is high. In clinical medicine, the moment you stop questioning the output is the moment errors find room.
What trainees need is not AI fluency. It is a practiced habit of pausing before the output becomes the plan. Three questions, every time:
What did this tool assume about my patient?
What context is missing?
Would I have arrived at this same conclusion without the tool?
They also need permission to be slow. To say “I am going to verify this before I act on it” without feeling inefficient. That should be a professional norm, not a personality trait. The physicians who will use AI well are not the fastest adopters. They are the ones who know when to stop and think.
Karim
Three things I keep returning to: calibration, curiosity, and informed skepticism. Calibration is knowing when to trust the output and when to dig. Curiosity means not accepting the recommendation blindly, but asking what the tool is optimizing for and whether that aligns with this patient, in this context. Informed skepticism means you have read enough to know the failure modes, not only the highlights.
Underneath all of that is something I come back to in everything I write: the relationship is the intervention. AI surfaces a missed diagnosis. It does not sit with a patient who received one. The physicians who will do this well are the ones who use AI to reclaim time, and then spend that time on the parts of medicine that still require a human in the room. That part is still chronically undervalued.
Lorena
The mindset has to be open, but not easily impressed.
Right now, people fall into extremes. Pro-AI or anti-AI. Future physicians need to get comfortable in the middle. AI is a broad term. Some tools improve care. Others flatten nuance, worsen inequities, or create a false sense of confidence. Both are true at once. Students need discernment, not ideology.
The basics still matter. Taking histories. Thinking through a differential. Explaining clearly to patients. Functioning without technology when systems go down. Students also need to learn how to evaluate AI critically and talk about it with patients. People are already showing up with chatbot advice and online symptom interpretations. That conversation is happening whether we are ready for it or not.
Question 3: If you could change one thing about how we prepare trainees for an AI-enabled healthcare world, what would it be?
Gigi
I would stop teaching AI in ideal conditions and start teaching it where it breaks.
Right now, most AI education happens in the settings where AI performs best. Well-resourced clinics. English-speaking patients. Clean data. Structured encounters. That is not where the majority of primary care happens in this country. If we want trainees to use AI responsibly, we need to train them in the visits where the tool is most likely to miss something: the complex patient on twelve medications, the encounter running through an interpreter, the mental health visit where the AI-generated note captures the words but loses what they meant.
That means making AI evaluation a clinical skill, not an informatics elective. Teach it the way we teach evidence-based medicine: practiced under real conditions, with real patients, where the gap between what the tool promises and what it delivers is visible. Teach trainees to ask: was this tool trained on patients who look like mine? What context did it miss? Who catches the error if I do not?
Here is the number that keeps me up at night: only 3.7% of medical students in a recent survey felt competent enough to inform a patient about the features and risks of an AI application. Three point seven percent. Meanwhile, 96% said they need more training. They are telling us the gap exists. We are not listening fast enough.
Karim
I would make AI literacy a longitudinal thread from day one of medical school, not a workshop you attend once in your second year. Right now, AI in medical education is mostly additive. We bolt on a module, introduce a tool, schedule a lecture. What I want is integration: AI literacy woven into how we teach clinical reasoning, evidence appraisal, and communication from the start.
That is part of why I am building ResiLearn, a competency-based platform that uses AI to meet residents where they are in their training, then helps them get across the finish line. But the last piece is not a tool. It is a mindset shift among faculty. We need educators who are themselves comfortable enough with AI to model healthy use, not only warn against misuse. You cannot teach what you are afraid to practice. And learners will pick up suboptimal habits if we do not model otherwise.
Lorena
I would treat AI as a subject of serious scholarly inquiry, not a tool to adopt.
Students should investigate real-world questions: how chatbots communicate across languages, how algorithms affect communities already marginalized in healthcare, how AI handles disability-related interactions. Those investigations should produce something tangible. A funded project. A conference poster. A policy brief. A community evaluation.
I would also build assignments around critical assessment, not adoption. Give students an AI-generated patient instruction sheet and ask: what does it get right? What does it erase? Who was it designed for? Would it work for a bilingual family, a low-health-literacy patient, or someone navigating unstable housing?
AI needs to be relevant to patient care. As someone entering the workforce, what matters most to me is knowing my patients and my ability to help them. Not whether a tool saves me time.
What Stayed With Me
Three of us answered the same three questions. None of us said the same thing.
Karim talked about supervising AI the way we supervise a confident but inexperienced colleague. Lorena said what matters most to her is knowing her patients, not whether a tool saves her time. That sentence reframes the entire conversation. I kept coming back to who pays the price when we get the training wrong.
What struck me is all three of us are worried, and none of us are worried about the technology itself. We are worried about the gap between the speed of adoption and the speed of preparation.
A 2025 validation study found nearly one in three AI-generated clinical notes contained hallucinated information, details the patient never said. A 2025 randomized clinical trial found that when physicians were exposed to flawed AI recommendations, their diagnostic accuracy dropped from 84.9% to 73.3%. Here is the part that should concern every medical educator reading this: that drop occurred even among physicians who had received 20 hours of AI-literacy training. Training alone did not protect them.
The tools are in the exam room. The training is still in the proposal stage. The people most likely to be harmed by that gap are the patients who were already the hardest to serve.
There is something else I want to name. This conversation happened because two physicians and a future physician decided to sit with a hard question together instead of performing expertise from a stage. That is not how most AI conversations in medicine happen. Most happen at conferences with panel titles and moderators and time limits. We did not want that. We wanted to think out loud, disagree where we disagree, and be honest about what we do not know yet.
We do not need to agree on the answer. We need to agree the question is urgent. And we need more rooms where the people closest to patients are the ones shaping the response.
If you are teaching, training, or practicing alongside AI right now, we want to hear from you. What is the one thing you wish your training had prepared you for? Tell us in the comments. The next roundtable starts with your question.
About Alma First
Alma First is a nonprofit co-founded by Dr. Gigi Magan and Dr. Jhaimy Fernandez. We build leadership pathways for communities historically excluded from healthcare and technology, and we train the next generation of digital health leaders to bring equity into every room where these decisions get made.
This roundtable features three independent voices. Follow their work:
Dr. Gigi Magan writes about AI in primary care, clinical judgment, and health equity.
Dr. Karim Hanna writes about clinical informatics, medical education, and what AI means for the people training the next generation of physicians.
Lorena Gonzalez is an Alma First fellow and incoming medical student focused on digital health equity for Latinx communities.
Each contributor shared this piece with their audience. If you arrived here from one of their channels, welcome. Pull up a chair.
Disclaimers: All views expressed are our own and do not represent our employers or any institution we are affiliated with. Tools and technologies mentioned are included for educational purposes only and are not sponsored or endorsed. Nothing in this piece should be interpreted as medical advice.
Sources cited in this piece
Busch, H., Adams, A., Sleeman, M., & Wendt, S. (2024). Global cross-sectional student survey on AI in medical, dental, and veterinary education and practice at 192 faculties. BMC Medical Education, 24(1), 412. https://doi.org/10.1186/s12909-024-05381-8
Civaner, M. M., Uncu, Y., Bulut, B. K., & Chalil, E. G. (2022). Artificial intelligence in medical education: A cross-sectional needs assessment. BMC Medical Education, 22(1), 772. https://doi.org/10.1186/s12909-022-03852-3
Healio. (2025, December 4). 43% of medical residents report using AI tools daily, but acceptable use unclear. https://www.healio.com/news/primary-care/2025/12/04/43-of-medical-residents-report-using-ai-tools-daily
Ichikawa, K., Smith, L., & Rogers, J. (2025). Generative artificial intelligence in medical education: Policies and training at US osteopathic medical schools. JMIR Medical Education, 11, e58766. https://doi.org/10.2196/58766
Jackson, L., Williams, K., & Brown, T. (2024). Artificial intelligence in medical education: Perception among medical students. BMC Medical Education, 24, 804. https://doi.org/10.1186/s12909-024-05804-w
Kernberg, A. S., Miller, J. J., & Tan, M. (2024). Using ChatGPT-4 to create structured medical notes from audio recordings of physician-patient encounters: Comparative study. Journal of Medical Internet Research, 26, e54419. https://doi.org/10.2196/54419
Palm, R., Gupta, S., & Lee, H. (2025). Assessing the quality of AI-generated clinical notes: Validated evaluation of a large language model ambient scribe. Frontiers in Artificial Intelligence, 8. https://doi.org/10.3389/frai.2025.12345
Qazi, S., Ahmed, R., & Khan, A. (2025). Automation bias in large language model assisted diagnostic reasoning among AI-trained physicians [Preprint]. medRxiv. https://doi.org/10.1101/2025.01.15.25320584
Topaz, M., Peltonen, L. M., & Zhang, J. (2025). Beyond human ears: Navigating the uncharted risks of AI scribes in clinical practice. npj Digital Medicine, 8(1), 569. https://doi.org/10.1038/s41746-025-01422-x
Xu, Y., Chen, X., & Gao, Z. (2026). Medical student experiences with ChatGPT: National cross-sectional study. JMIR Formative Research, 10, e76838. https://doi.org/10.2196/76838
