Most medical students already use generative AI to study, summarize and practice diagnosis, often before their schools have taught or regulated it. Curricula are catching up quickly, but the bigger problem is assessment. Exams and assignments that AI can complete no longer measure what a future doctor can do, so schools need to test reasoning, not recall.
In early 2023 a team of researchers reported in PLOS Digital Health that ChatGPT, a general chatbot with no medical training, had performed at or near the passing threshold on all three steps of the United States Medical Licensing Examination. The finding was widely reported as a curiosity. For medical students, it was a practical announcement: the tool that could answer exam questions was free, available at 2 a.m., and never impatient.
Students did not wait for permission. By the time many schools had written their first AI policy, students were already using chatbots to summarize lectures, generate practice questions, explain physiology in plain language and draft differential diagnoses for cases.
The curriculum is now moving. But the argument of this piece is that curriculum is the easier half. The harder half is assessment. If an essay, a case write-up or a multiple-choice bank can be completed well by a model, it no longer tells a school whether its student can think like a doctor. Medical education has to change what it measures, not only what it teaches.
How many medical students use AI?
The data are patchy but point in one direction.
A 2025 survey of students at six medical schools in Ontario, published in the Journal of Medical Education and Curricular Development, found that 78.9% of respondents reported using generative AI, and 53% used it at least weekly. The most common uses were reviewing medical content, learning concepts and creating study guides. Nearly half reported using it to generate differential diagnoses. The survey had a low response rate, 167 of about 3,800 eligible students, so enthusiastic users may be over-represented. It is a signal, not a census.
Among qualified doctors, better data exist. The American Medical Association’s physician survey, fielded in early 2026, found that more than 80% of responding physicians used AI professionally, double the 2023 share. It also found that concern about losing clinical skills to AI was higher among physicians in the first decade of their careers, the group closest to medical school.
The consistent picture across surveys is that use is high, frequent and largely self-taught, and that students’ main uses are learning tasks that overlap heavily with what schools assess.
Are medical schools teaching AI?
Increasingly, yes, and quickly.
The Association of American Medical Colleges reported in June 2025 that 140 US and Canadian medical schools covered AI in their curricula in the 2024 academic year, up from 88 the year before, out of 182 schools that responded to its curriculum survey. Of those 140, 107 included AI in required rather than elective teaching. The most common topics were ethical use and basic AI concepts.
That is a fast shift for medical education, which usually changes over years. But most of it is teaching about AI: what it is, where it fails, how to use it responsibly. That is necessary. It is not the same as redesigning the rest of the curriculum for a world in which every student carries a capable assistant.
The distinction matters because students are not waiting for the AI module to change how they study. They are using AI across every course now. A lecture on AI ethics in second year does not address what happens when first-year students have a chatbot write their reflective essays.
Why is assessment the real problem?
Medical assessment has long relied on three kinds of task: written assignments, knowledge tests and supervised clinical performance. AI hits the first two much harder than the third.
Written assignments, such as essays, case reports and reflective pieces, can now be drafted by AI to a passable standard. Detection tools are unreliable, and policing them turns educators into investigators. If a school keeps grading unsupervised written work as if it reflected a student’s unaided thinking, it will certify skills students may not have.
Knowledge tests taken in secure conditions still work for measuring recall. But recall is becoming less central to practice when doctors can look things up instantly. The skill that matters more is judgment: knowing which question to ask, which answer to distrust and what to do when the information is incomplete.
Supervised clinical assessment, in which a student takes a history, examines a patient, explains a plan and is observed doing it, is the hardest for AI to fake. It is also the most expensive to run, which is why many schools rely on it less than they would like.
There is also an equity problem hiding here. Students with paid subscriptions to stronger models, or with more confidence in using them, gain an advantage on unsupervised written work that has nothing to do with clinical aptitude. A grading system that rewards polished AI-assisted output quietly rewards access, not ability.
So AI pushes assessment toward the expensive end. Schools that do not fund that shift will end up measuring something increasingly disconnected from competence.
What should medical schools assess instead?
Several approaches are emerging, and they share a principle: assess the reasoning process, not only the product.
Observed reasoning. Oral case presentations, viva-style questioning and observed structured clinical examinations reveal whether a student can reason under uncertainty without assistance. They are costly but hard to outsource.
AI-inclusive tasks. Instead of banning AI in written work, some educators assign tasks that require its use and assess the student’s critique: give the student an AI-generated management plan with errors and ask them to find and correct them. This tests the skill doctors will actually need, supervising a machine.
Process evidence. Requiring drafts, prompts, reasoning notes or short oral defenses of written work makes the student’s contribution visible.
Unassisted foundations. For core skills, students should demonstrate competence without assistance first. We have argued that residents should learn to work without AI first, and the same logic applies earlier. A student who has never built a differential diagnosis unaided cannot judge whether the AI’s version is any good.
A worked example: redesigning one assignment
Consider a hypothetical second-year assignment: a 1,500-word written case analysis of a patient with chest pain, graded on accuracy and completeness.
In its current form, a student can paste the case into a chatbot, edit the output for twenty minutes and submit a strong answer. The grade measures the student’s editing and the model’s knowledge.
Redesigned, the assignment has three parts. First, in a supervised thirty-minute session with no devices, the student writes a short differential diagnosis and initial plan. Second, the student is given an AI-generated analysis of the same case, which contains two deliberate errors, one a dangerous omission, and must identify and correct them, with reasoning. Third, a ten-minute oral discussion with a tutor explores why the student prioritized as they did.
The redesigned version takes more faculty time. It also measures three things the original did not: unaided reasoning, the ability to catch a machine’s mistakes, and the ability to explain a decision to another clinician. Those are closer to what the student will do as a doctor than a polished essay.
What is the strongest case for letting students use AI freely?
The strongest objection to this approach is that it looks backward. Doctors will practice with AI, so students should learn with it from the start. Insisting on unassisted performance may be like insisting that students learn to calculate drug doses by slide rule. Restrictions also push AI use underground, where it is less safe and less discussed.
There is real force here. Banning AI is unenforceable and probably counterproductive, and students who never learn to use it well will be less prepared for practice.
But the objection conflates two things: using AI and depending on it. A doctor needs enough independent knowledge and reasoning to recognize when an AI output is wrong, incomplete or dangerous. That capacity is built by practice without the tool, and it cannot be assessed if every task allows the tool. The goal is not an AI-free education. It is an education that can tell which skills the student has and which belong to the software.
What this means for medical schools, students and regulators
For deans and curriculum leaders, the priority is an assessment audit. Go through every graded task and ask whether an AI could complete it well without the student understanding the material. Where the answer is yes, redesign the task or change what it counts for. Budget for more observed assessment, because that is where the cost of reliable measurement now sits.
For faculty, AI-inclusive tasks that require critique are among the most useful teaching tools available. They build exactly the supervisory skill students will need and are hard to game.
For students, the practical advice is to use AI to learn, not to replace learning. Build the differential yourself, then compare. Use the chatbot to quiz you, not to answer for you. The skills you do not build now will be hard to build later, and patients will rely on them.
For accreditors and licensing bodies, the question is whether current standards and examinations still measure competence when AI is ubiquitous. Continuing education has long paid for attendance rather than change; undergraduate assessment should not make the same mistake by grading outputs a machine produced.
Frequently asked questions
Do medical students use ChatGPT?
Many do. A 2025 survey of Ontario medical students found 78.9% of respondents used generative AI and 53% used it weekly, mainly to review content, learn concepts and create study materials. The survey’s low response rate means it may overstate use, but other surveys also report high adoption among medical students.
Is it cheating for medical students to use AI?
It depends on the school’s policy and the task. Many schools permit AI for studying but restrict it in graded work unless the assignment allows it. Students should check their institution’s rules and disclose use where required. The larger question for schools is whether their assessments still measure what students can do unaided.
How many medical schools teach AI?
According to the Association of American Medical Colleges, 140 US and Canadian medical schools covered AI in the 2024 academic year, up from 88 in 2023, among schools responding to its curriculum survey. Of those, 107 included AI in required teaching, most commonly ethics and basic concepts.
Can AI pass medical licensing exams?
Studies suggest general-purpose AI models can perform at or near passing levels on multiple-choice licensing questions. A 2023 PLOS Digital Health study found ChatGPT performed at or near the passing threshold on all three USMLE steps. That says more about the limits of multiple-choice exams than about AI’s readiness to practice medicine.
How should medical students use AI safely?
Use it as a study partner: to explain concepts, generate practice questions and compare against your own reasoning. Verify its claims against trusted sources, because models can produce confident errors. Never enter identifiable patient information into public tools, and follow your school’s and clinical placement’s policies on AI use.
Students have already answered the question of whether medicine will use AI. What schools have not yet answered is how they will know which of their graduates can practice when the machine is wrong.
If you teach medical students, which of your assessments would an AI pass today, and what are you replacing it with?
