AI in medical education: what works, and what to be careful about

Two years ago the question was whether artificial intelligence belonged in medical education at all. Today the question is which uses are worth your time and which are quietly making you a worse doctor. Having built an AI-based platform ourselves, we have strong opinions on both.

Where it genuinely helps

Practice at scale. The oldest bottleneck in clinical training is the number of patients a student can see and the number of teachers who can watch. A simulated patient that answers questions in natural language, reacts to your management and never gets tired removes that bottleneck. You can take fifty focused histories in a week instead of five.

Immediate, specific feedback. Most feedback in medical school arrives late and vague ("read more on this"). A well-designed system can tell you, the moment you submit a plan, which red flag you did not ask about and which investigation you ordered without a reason. Feedback that is immediate and specific is what turns repetition into learning.

Adaptive questioning. Question banks that track what you get wrong and resurface it — combined with spaced repetition — are an old idea that AI makes cheaper to personalise.

Reading guidelines faster. Summaries with citations can get you to the relevant paragraph of a 200-page guideline in a minute. The operative words are with citations.

Where you should be careful

Fabricated references and confident errors. Large language models produce fluent text, and fluency is not accuracy. They can invent a trial, a dose or a guideline recommendation, and they do so in the same confident tone as when they are right. Any clinical statement from an AI tool that you intend to act on, write in an exam or tell a patient must be checked against the primary source.

Patient data. Never paste a real patient's details into a general-purpose chatbot. It is a privacy breach in most jurisdictions and an ethical one everywhere.

Deskilling. If the tool produces the differential diagnosis for you, you do not learn to produce it. Use AI to rehearse a skill with feedback, not to perform the skill for you. The right question to ask any tool is: "After I close this, can I do this on my own?"

Bias. Models learn from the data they are trained on; patients who are under-represented in that data — including many populations in Central Asia — are the ones for whom the answers are least reliable.

What the WHO says

The World Health Organization's guidance on the ethics and governance of AI for health sets out six principles: protect human autonomy; promote human well-being, safety and the public interest; ensure transparency, explainability and intelligibility; foster responsibility and accountability; ensure inclusiveness and equity; and promote AI that is responsive and sustainable ^[1]. Its 2024 guidance on large multimodal models adds specific recommendations for their use in education and clinical support, including the need for human oversight and validation before deployment ^[2]. The AMEE guide on the ethical use of AI in health professions education translates these principles into practical rules for students and teachers ^[3].

A practical rule for students

Use AI where a tireless practice partner helps — histories, examinations, question drills, explaining a concept three different ways — and keep your own reasoning in the loop. Treat every factual claim as a lead to verify, not a fact to memorise. The clinicians who benefit most from these tools will be the ones who know enough to notice when the tool is wrong.