AI Virtual Patients: How Medical Students Can Practise History-Taking and OSCEs
What AI virtual patients are, which tools medical students use, what they train well and badly, and how Malaysian faculties can add them to OSCE preparation.
An AI virtual patient is a simulated patient driven by a language model: the student asks questions in their own words, the patient answers in character, and the tool gives feedback on the history and the reasoning. It is good for extra practice in history-taking and differential diagnosis. It does not train physical examination, practical skills or the human side of a real consultation, and it does not replace a supervised OSCE.
Part of the guide: AI in healthcare and medicine

On this page
A medical student meets a limited number of patients before finals. A timetable cannot add more, but software can. AI virtual patients let a student take a history at midnight, get it wrong, and try again. The question for faculties is not whether students will use them. Many already do. The question is how to make that practice count.
Checked on 7 November 2026. This article is about medical education. It is not medical advice and not a product review.
What an AI virtual patient does
Older virtual patients worked from menus: the student picked the next question from a list. AI virtual patients are different in three ways.
- Free conversation. The learner asks in their own words, in the order they choose, the way they would on the ward.
- A hidden script. The case sits behind the conversation. The patient reveals the chest pain's character only when asked about it.
- Feedback. Many tools mark which important questions were missed, compare the learner's differential with the case, and suggest what to read.

Tools worth knowing
These are public examples we show in our teaching. We are not affiliated with any of them, and features change, so check the current version before you recommend one to students.
| Tool | What it offers | Good for |
|---|---|---|
| Oscer | AI virtual patients and clinical reasoning practice | History-taking and differential diagnosis before OSCEs |
| Body Interact | Screen-based patient simulation with AI-driven dialogue | Managing a patient whose condition changes over time |
| Glass Health | AI drafts differentials and plans grounded in literature | Comparing your reasoning with an AI's, then finding where it is wrong |
| A general model (ChatGPT, Claude, Gemini) | Role-play from a case script you write | Faculty-built cases matched to your own curriculum |
More examples, including anatomy and 3D projects, are in the showcase on our AI in healthcare and medicine page.
What they train well, and what they do not
Trains well
- Structured history-taking and not forgetting the systemic review
- Building and narrowing a differential diagnosis
- Choosing initial investigations
- Volume: many more cases than clinical placements give
Trains poorly or not at all
- Physical examination and practical procedures
- Reading body language, silence and distress
- Communication under real time pressure and with an examiner watching
- Local context: language mix, family involvement, cost questions in Malaysian practice
As our article on simulation-based training notes, a 2026 systematic review of language-model virtual patients found that most systems simulate internal medicine and mental health cases, with few covering multimorbidity or rare conditions. Students who practise only on AI cases get a narrower case mix than they think.

How a faculty can add them without losing control
Students will find these tools on their own. A faculty that guides the use gets better practice and fewer surprises.
- Pick the purpose. Extra history-taking practice before OSCEs is a sensible start. Assessment is not.
- Approve the cases. Write or approve a small library of constructed cases that match your curriculum, including common Malaysian presentations such as dengue, diabetes complications and tuberculosis.
- Ask for the transcript. A student who brings the conversation to a tutorial gives the tutor something concrete to correct.
- Debrief, briefly. Ten minutes on "what did you miss and why" turns practice into learning.
- Keep humans in the OSCE. Simulated patients and examiners stay the standard for summative assessment.
Our faculty article on simulation-based clinical training with AI covers scenario design and debriefing in more depth, and AI for medical students in Malaysia covers how students can use AI without losing skills.
A simple prompt for a faculty-built virtual patient
If you want to try the idea before buying anything, a general model can play a patient from a case you write. Give it four parts: the role ("you are a 58-year-old man in the emergency department"), the hidden case details, the rule ("reveal details only when asked directly, answer in plain language, stay in character"), and the end condition ("when the student says 'end of history', give feedback on missed questions"). Test it yourself twice before students use it, and use an institution-approved account.
Teach your faculty to build and use them
Designing cases, simulation scenarios and virtual patient exercises with faculty validation is part of our course AI in Medical & Clinical Education, led by two emergency physicians. It runs in-house for medical schools and hospitals, and registered employers can claim it from HRD Corp.
Frequently asked questions
What is an AI virtual patient?
A simulated patient powered by a large language model. Instead of choosing from fixed menu questions, the learner talks or types freely, and the virtual patient answers from a hidden case script. Many tools add feedback on missed questions and on the learner's differential diagnosis.
Are AI virtual patients good for OSCE preparation?
They help with the history-taking and clinical reasoning parts of an OSCE and give students many more practice cases than a timetable allows. They do not prepare students for examination skills, procedures or being observed by an examiner.
Can students build their own AI virtual patient with ChatGPT?
Yes, with a prompt that sets a case script and tells the model to stay in role and only reveal details when asked. It works for practice, but the case should come from a faculty-approved scenario, and the student should check the feedback against their teaching.
Do AI virtual patients use real patient data?
Good tools use constructed cases. Faculties building their own should do the same: real cases, even lightly changed, can identify a patient, and the MMC guideline forbids sharing patient data on an AI system without consent.