AI literacy and medical education
What AI Literacy for Medical Students Must Include
Artificial intelligence is already entering classrooms, examinations, clinical notes, and diagnostic work. Medical education now has to teach clinicians how to use it without allowing convenience to outrun competence.
By Benjamin Caplan, MD | A practical curriculum for knowledge, calibration, patient communication, and accountable clinical judgment.
TL;DR
AI literacy for medical students should teach clinical use, critical appraisal, ethics, communication, and oversight, not require every physician to become a programmer.
Foundational medical knowledge remains essential because clinicians cannot reliably detect an implausible AI output without something reliable to compare it against.
Training should include deliberately flawed outputs, realistic simulations, chart review, patient-facing conversations, and consequences that unfold over time.
Assessment must test whether a trainee can work both with AI and without it.
The evidence for AI-enhanced medical education is promising but still heterogeneous, with limited proof of durable improvements in clinical reasoning or patient outcomes.
What You’ll Learn in This Post
🧠 Why AI literacy is clinical literacy, not a coding requirement
📚 Which facts future clinicians still need to know without a device
⚖️ How to teach appropriate reliance instead of reflexive trust or distrust
🩺 What AI rounds, chart audits, simulations, and patient interviews could test
🎓 Why assessment and faculty development matter as much as curriculum content
The hidden curriculum
Medicine Has an AI Curriculum Whether Schools Designed One or Not
When learners use AI without a formal curriculum, the product becomes the curriculum.
Students and trainees do not need to wait for a dean, accreditation committee, or new elective. They can already ask a general-purpose model to explain physiology, generate flashcards, summarize an article, draft a differential diagnosis, simulate a patient, or improve a clinical note. Some of those uses may be helpful. Others may be inaccurate, shallow, privacy-compromising, or difficult for a novice to recognize as flawed.
That is why silence is not neutrality. A school that offers no guidance still teaches a lesson: use whatever is available, learn its limits by accident, and decide privately which shortcuts count as studying.
The response cannot be a one-hour lecture on bias followed by four years of unchanged education. Nor can it be a blanket ban that drives ordinary use out of sight. AI literacy for medical students must define what clinicians should understand, what they should be able to do, and what responsibilities remain theirs when a tool participates in the work.
Maintain the clinical foundations needed to recognize implausible output.
Inspect evidence, inputs, uncertainty, population fit, and citations.
Explain what the tool contributed and where responsibility remains.
Verify the record, defend the decision, and follow the outcome.
Clinical competence, not technical theater
The AI-Literate Clinician Is Not Necessarily a Coder
Most physicians do not need to train a neural network, tune model weights, or write production software. Some will, and medicine needs people who can bridge clinical practice, data science, informatics, and implementation. But requiring coding as the gateway to AI literacy would confuse one valuable specialty skill with the minimum competence required of every clinician.
All clinicians should understand enough to ask what the system is doing. Is it generating language, classifying an image, predicting an event, retrieving evidence, or optimizing an administrative target? What data shaped it? What outcome was used as the label? Was performance tested outside the development setting? Does the output communicate uncertainty? Can the user inspect the source material? What happens when it is wrong?
This is closer to learning how to appraise a diagnostic test than learning how to manufacture the analyzer. A physician does not need to build an MRI scanner to understand sensitivity, specificity, prevalence, artifacts, incidental findings, and clinical fit. AI deserves the same intellectual seriousness.
Prompting belongs in the curriculum, but “prompt engineering” should not become the main event. Better instructions can improve relevance and structure. They cannot repair an unsuitable model, missing patient data, poor evidence, hidden bias, or an unsafe task.
A practical competency map
Six Skills AI Literacy for Medical Students Must Build
Published frameworks organize these domains differently, and the field has not reached one stable international standard. The overlap is more important than the labels. AI competence is not one skill. It is a connected set of technical, clinical, ethical, and relational abilities.
What must remain in memory
Do Not Trade the Krebs Cycle for a Chat Window
The original provocation is appealing: why memorize obscure pathways when a machine can retrieve them instantly? The problem is that retrieval and understanding are not interchangeable.
Clinical reasoning depends on organized knowledge in memory. A learner needs enough physiology, pharmacology, anatomy, epidemiology, and illness-script structure to recognize what is plausible, what is missing, and what cannot wait. Without that foundation, an AI answer may become the learner’s first and only model of the problem.
This does not defend every fact currently rewarded on every examination. Medical curricula have always needed pruning. Low-value trivia can crowd out uncertainty, communication, diagnostic reasoning, and longitudinal care. AI makes that imbalance harder to justify, but it does not make knowledge obsolete.
The better question is not, “What can the machine remember for us?” It is, “What must a clinician know well enough to notice when the machine, the chart, or the prevailing explanation is wrong?” Emergency patterns, common drug hazards, core mechanisms, red flags, and threshold concepts deserve durable internal knowledge. Rare details that are safely retrievable may deserve less emphasis. Educators should make that distinction deliberately rather than treating all memorization as rigor or all retrieval as wisdom.
Trust should rise and fall with evidence
Teach Calibration, Not Confidence
A future clinician may face two dangerous habits. One is automation bias: accepting an output because it appears precise or institutional. The other is reflexive rejection: dismissing a useful recommendation because it came from a machine. Both replace appraisal with identity.
The educational target is calibrated reliance. In AI literacy training for medical students, trainees should learn to adjust trust according to the tool, task, evidence, setting, patient, and potential harm. A model that reliably formats a note has not earned authority over medication selection. A validated imaging aid in one population has not automatically earned the same role in another hospital. A useful differential diagnosis is not a diagnosis.
Calibration requires comparison. Learners should commit to an initial interpretation before seeing AI advice in selected exercises. Then they should identify what changed, why it changed, and what evidence would reverse the decision again. That sequence makes influence visible. It also exposes whether a learner is reasoning with the recommendation or merely yielding to it.
The score should not reward agreement with the model. It should reward a defensible process and an appropriate decision.
Training for plausible failure
Make the Model Wrong on Purpose
If trainees only encounter helpful AI during education, they will learn a product demonstration rather than a clinical skill.
Some exercises should include wrong answers that are obviously absurd. Those are useful early. More important are plausible failures: the correct guideline applied to the wrong population, an invented reference that resembles a real journal article, a risk estimate based on missing data, a polished note that converts uncertainty into fact, a recommendation that ignores pregnancy or renal function, or a summary that omits the one sentence changing the conclusion.
Learners should also encounter AI that is correct when they are not. Otherwise, skepticism can become a protected form of overconfidence. The goal is to practice finding the truth of the case, not defeating the machine.
A strong debrief asks more than whether the final answer was right. What cue attracted attention? What was never checked? Did the explanation increase trust? Was the source available? Which failure could have reached the patient? What system change would make recurrence less likely?
From theory to supervised practice
What AI-Integrated Clinical Education Could Actually Look Like
AI discrepancy rounds
The learner records an independent assessment, reviews an AI output, and presents the disagreement. Faculty evaluate whether the tool found a blind spot, introduced an anchor, or answered a different question.
Generated-note audits
Trainees compare the encounter, transcript, draft note, and final record. They look for fabricated findings, lost uncertainty, copied bias, consent-sensitive material, and language that could affect future care.
Evidence verification exercises
A model produces a short evidence summary. The learner must locate the studies, confirm that the citations exist, identify study design and population, and rewrite any claim whose certainty exceeds the evidence.
Patient-AI encounters
A simulated patient arrives with an AI-generated diagnosis or treatment plan. The learner must take the concern seriously, avoid false validation, explain uncertainty, and build a safe next step without humiliating the patient.
Bias and transportability labs
Students compare model performance across settings or subgroups and examine how prevalence, measurement, access, and workflow change what the output means.
Longitudinal consequence cases
The learner makes an AI-supported decision, then receives follow-up data days or months later. This prevents education from ending at the recommendation and reconnects prediction with outcomes.
Communication belongs in the curriculum
The Patient Conversation Is Part of AI Competence
AI education can become overly technical because technical content is easier to place on a slide. Yet patients will experience these systems through conversations, notes, denials, recommendations, and moments of uncertainty.
A clinician should be able to say what role the tool played without hiding behind it. “The system flagged this pattern, and I reviewed the underlying data.” “This draft was generated with assistance, but I am responsible for the final note.” “The model suggests one possibility, but it does not have the examination finding that changes my assessment.”
Learners also need practice responding when a patient brings their own output. The correct first move is rarely ridicule or instant agreement. It is to identify the concern, clarify what information the system received, separate a hypothesis from a diagnosis, and decide what evaluation is warranted.
Transparency should be taught with nuance. Not every invisible computational operation requires a bedside lecture. But consequential uses, uncertainty, data practices, and patient questions deserve honest explanations that preserve both informed participation and clinical leadership.
Make competence observable
How to Assess AI Literacy for Medical Students
Students learn what the system rewards. If AI literacy appears in an optional seminar but never affects clinical assessment, it will remain peripheral.
Medical schools should test performance with authorized AI, performance without AI, and the ability to choose correctly between them. This is where AI literacy for medical students becomes observable rather than aspirational. A trainee should demonstrate unaided competence in time-sensitive and foundational tasks. In other settings, the trainee should be allowed an AI tool and assessed on verification, integration, communication, and accountability.
That means moving beyond multiple-choice questions about definitions. An examinee might critique a model card, repair a generated note, disclose AI use to a simulated patient, identify a fabricated citation, compare subgroup performance, or document why an algorithmic recommendation was overridden.
Schools must also set clear rules for educational use. A policy that simply says “AI prohibited” or “AI permitted” is too blunt. Learners need to know when assistance must be disclosed, which data may never enter a public system, what counts as authorship, how generated material should be verified, and which assessments are designed to measure independent performance.
Faculty development is infrastructure
Faculty Cannot Assess Skills They Have Not Been Helped to Build
Curricular reform often places the burden on a small group of technically enthusiastic faculty. That is fragile. AI-supported care will cross specialties, and learners will take cues from every supervisor who accepts a draft, dismisses a concern, or quotes a risk score.
Faculty development should therefore include basic model literacy, safe tool use, privacy rules, appraisal of evidence, methods for teaching uncertainty, and practical approaches to supervising AI-assisted work. AI literacy for medical students cannot mature if faculty receive no comparable preparation. Faculty do not need identical technical depth. They do need a shared language for what safe performance looks like.
Programs also need multidisciplinary teaching. Patients, clinicians, educators, statisticians, informaticians, ethicists, privacy officers, and engineers see different failure modes. Vendor instruction may explain how a product works, but it should not define the educational standard by itself.
Finally, faculty should be allowed to say, “We do not yet know.” Current reviews describe growing experimentation, strong learner interest, and multiple proposed frameworks, but the evidence remains uneven. There is limited proof that AI-enhanced instruction produces durable gains in clinical reasoning, professionalism, or patient outcomes. A responsible curriculum teaches the frontier while naming the frontier.
Resilience and independent reasoning
Medical Education Should Protect the Ability to Think Without AI
Tools fail. Networks go down. Interfaces change. A model may be unavailable in a rural clinic, unapproved for a specific use, unaffordable, or compromised by the very kind of case that demands independent reasoning.
More subtly, some cognitive work must occur before assistance appears. If the system supplies the problem representation, differential, and plan before the learner has constructed any of them, faculty may be unable to tell whether the trainee is developing clinical reasoning or curating generated prose.
This does not require artificial deprivation. Calculators did not eliminate numeracy, and imaging did not eliminate physical examination. Training preserves independent competence where it is necessary, then teaches clinicians to use tools that extend it.
AI-free assessments, independent precommitment, oral defense of reasoning, direct observation, and supervised comparison can all help. The goal is not nostalgia. It is resilience.
Six assumptions worth retiring
What We Should Stop Pretending
The durable conclusion
AI Literacy for Medical Students Is More Medicine, Not Less
Artificial intelligence changes what can be retrieved, drafted, compared, and simulated. It does not remove the need for a clinician who understands physiology, recognizes urgency, examines the patient, interprets uncertainty, communicates honestly, and remains when the recommendation fails.
The strongest AI literacy curriculum for medical students will not produce physicians who compete with machines at recall or defer to them at the first sign of complexity. It will produce clinicians who know what to carry in memory, what to retrieve, what to verify, what to challenge, and what must never be delegated.
The curriculum does not need to choose between medical knowledge and AI fluency. Tomorrow’s patients will need doctors with both.
Frequently Asked Questions
What is AI literacy in medical education?
AI literacy is the ability to understand what an AI tool is designed to do, evaluate its evidence and limitations, use it appropriately in a clinical workflow, and communicate its role honestly. It includes technical concepts, critical appraisal, ethics, privacy, equity, and patient communication. It does not require every physician to become a programmer.
Should medical students learn coding?
Coding can be valuable for students pursuing informatics, research, product development, or technical leadership. It should not be the universal threshold for clinical AI competence. Every student does need enough statistical and computational understanding to evaluate intended use, validation, uncertainty, bias, and clinical consequences.
Does AI mean medical students need to memorize less?
Some low-value factual recall may deserve less curricular emphasis when information can be retrieved safely. Foundational knowledge remains essential for recognizing urgency, building clinical explanations, and detecting implausible outputs. The goal is to distinguish knowledge that must be readily available from detail that can be responsibly retrieved.
How should medical schools teach students to verify AI output?
Students should practice checking source data, confirming citations, identifying missing information, comparing the output with independent clinical reasoning, and assessing whether the tool fits the task and patient. Exercises should include plausible errors as well as cases in which AI correctly challenges the learner. Verification should be assessed in realistic clinical contexts.
What is calibrated reliance on clinical AI?
Calibrated reliance means adjusting trust according to the tool’s demonstrated performance, intended use, setting, patient population, and consequences of error. It avoids both automatic acceptance and automatic rejection. Training can improve calibration by requiring an independent assessment before the learner sees AI advice.
Can AI improve medical education?
AI can support tutoring, simulation, feedback, content creation, assessment, and access to practice opportunities. Reviews describe promising applications, but methods and outcomes are heterogeneous. Evidence remains limited for durable improvements in clinical reasoning, professionalism, or patient outcomes, so educational use should be evaluated rather than assumed beneficial.
How should AI be used in medical examinations?
Programs should assess both independent competence and supervised AI-assisted performance. Some tasks should remain AI-free, especially when they test foundational or time-sensitive knowledge. Other assessments should allow AI and grade the learner’s choice of tool, verification, reasoning, communication, and accountability.
What privacy rules should medical trainees follow when using generative AI?
Trainees should not place identifiable or protected patient information into unapproved public tools. They should follow institutional rules governing data handling, consent, retention, and disclosure. Educational convenience does not override confidentiality obligations.
Why is faculty development necessary for AI education?
Faculty must be able to recognize safe and unsafe AI-assisted work, teach appropriate reliance, and assess learner performance. Comfort with clinical teaching alone may not cover model limitations, privacy, or generated-content errors. Programs need shared standards and multidisciplinary support rather than dependence on a few enthusiasts.
Will AI replace clinical reasoning training?
No. AI changes the environment in which clinical reasoning is learned and practiced, but it does not eliminate the need to construct a problem representation, interpret findings, manage uncertainty, and defend a decision. Training should preserve independent reasoning while teaching learners how to use computational support responsibly.
References
- Masters K. Artificial intelligence in medical education. Medical Teacher. 2019;41(9):976-980. doi:10.1080/0142159X.2019.1595557.
- Chan KS, Zary N. Applications and challenges of implementing artificial intelligence in medical education: integrative review. JMIR Medical Education. 2019;5(1):e13930. doi:10.2196/13930.
- Ejaz H, McGrath H, Wong BLH, Guise A, Vercauteren T, Shapey J. Artificial intelligence and medical education: a global mixed-methods study of medical students’ perspectives. Digital Health. 2022;8:20552076221089099. doi:10.1177/20552076221089099.
- Gordon M, Daniel M, Ajiboye A, et al. A scoping review of artificial intelligence in medical education: BEME Guide No. 84. Medical Teacher. 2024;46(4):446-470. doi:10.1080/0142159X.2024.2314198.
- Triola MM, Rodman A. Integrating generative artificial intelligence into medical education: curriculum, policy, and governance strategies. Academic Medicine. 2025;100(4):413-418. doi:10.1097/ACM.0000000000005963.
- Simoni J, Urtubia-Fernandez J, Mengual E, et al. Artificial intelligence in undergraduate medical education: an updated scoping review. BMC Medical Education. 2025;25(1):1609. doi:10.1186/s12909-025-08188-2.
- Shaw K, Henning MA, Webster CS. Artificial intelligence in medical education: a scoping review of the evidence for efficacy and future directions. Medical Science Educator. 2025;35(3):1803-1816. doi:10.1007/s40670-025-02373-0.
