How to Use ChatGPT to Practise History Taking
Can you use ChatGPT to practise history taking?
Yes. A general chatbot can role-play a patient convincingly enough to rehearse the conversation parts of a history, from openings and open questions to ICE, red flags and summarising. It will play along, but it only knows what you tell it, so your setup decides how honest the practice feels.
In one study of 28 students and 826 question answer pairs, answers were drawn from the script 94.4% of the time when the question was covered, and 97.9% of answers were rated plausible. Use fictional cases only, never real patient details.
The simple loop that works tonight is four steps. Write a short case file, give the model a tight patient prompt, run the consultation by voice or text, then make it mark you. A later section is blunt about where it drifts.
What the research actually found
Tübingen 2024. Researchers built a chatbot on GPT-3.5 with an illness script plus behaviour rules, not the ChatGPT app, as reported in a Tübingen GPT-3.5 study. With 28 students and 826 question-answer pairs, 94.4% of answers used the script when the question matched; when it did not, 56.4% of 195 answers used invented information, yet 97.9% were rated plausible and usability was 77 out of 100.
Tübingen 2024, GPT-4 follow-up. Same group, purpose-built chatbot on GPT-4 with a written illness script, not the ChatGPT app, tested in a follow-up GPT-4 study. Across 106 conversations and 1,894 pairs, role-play was medically plausible in more than 99%; its feedback on history completeness agreed almost perfectly with a human rater, Cohen kappa 0.832, though agreement was weaker, kappa under 0.6, in 8 of 45 feedback categories.
South Africa 2026. Second-year students practised history taking with ChatGPT 3.5 and 4.0 using prompts built around Calgary-Cambridge, as reported in a South African pilot. Of 200 invited, 40 answered the survey; 78% felt better prepared for their OSCE and 75% rated the scenarios realistic, both self-report from a fifth of the class.
UK 2025, voice mode. In a UK voice-mode study, 27 students from three UK universities spoke to freely available ChatGPT-4o advanced voice mode playing the patient, using pre-written prompts. Self-reported confidence rose, for example breaking bad news from a median of 2 to 4 out of 5; no control group, small sample, self-assessment.
Step 1: write the case file before you open ChatGPT
Write the illness script first. Tübingen 2024 found that when questions fell outside the script, 56.4% of chatbot answers used invented information, so your guardrail is a written brief that covers what the patient knows and what they do not.
Use fictional cases only. Never include real patient details. Better still, get a friend to write it, or lift one from a textbook, so you do not already know the answer you are trying to uncover.
- Who the patient is, with a name, age, and job.
- The opening line, one sentence in their words, not yours.
- Presenting complaint and timeline, with clear onset and evolution.
- What they volunteer unprompted in the first minute.
- Facts they only give if asked, each paired with the exact question that unlocks it.
- Their ideas, concerns and expectations, phrased as they would say them.
- The answer you are aiming for, and the teaching point, kept in a separate note.
If you want to see what a complete one looks like, this is a worked chest pain history.
Step 2: the patient prompt to paste in
This is a fictional scenario. Do not use any real patient details.You are a patient, not a doctor.Stay in character the whole time.Case file, hidden from me: [paste case file here].Answer only what I ask, in a patient's everyday words.Keep answers short, 1 to 3 sentences.Never volunteer facts from the hidden list unless my question reaches them.If I ask something not in the case file, say you're not sure rather than inventing anything important.Never name a diagnosis or suggest tests.If I pause, wait. Do not take over.Don't give feedback until I type or say 'end of consultation'.
These guardrails exist because free chat often fills gaps. In Tübingen 2024, models invented answers to unscripted questions 56.4% of the time, so we tell it to keep replies short and to admit uncertainty rather than fabricate.
Voice brings its own quirks. A UK voice study in 2025 found ChatGPT frequently interrupted when students paused, ended scenarios early, or shifted into the physician role despite explicit instructions not to, so we hard block diagnosis talk, tests, takeover and midstream feedback.
If it drifts, re paste the rules and carry on. A quick reset is faster than arguing with a chatbot.
Step 3: run the consultation by voice or text
Use voice if you want the pressure of saying it out loud with no time to edit. A UK voice study in 2025 found ChatGPT-4o voice sometimes interrupted when students paused, which is a limitation, so make your turn-taking clear and finish questions cleanly.
- Introduce yourself, confirm their name and how they’d like to be addressed, and check you’ve got the right person for this fictional case.
- Open with a broad question, then let them talk. Don’t jump in until they stop.
- Explore the presenting complaint in depth, for example with SOCRATES where it fits the symptom.
- Cover ICE, and use plain words. Try this primer on asking about ideas and concerns.
- Screen for red flags that would change urgency or setting.
- Take the rest of the history, brisk and systematic: past medical history, drug and allergy history, family history, social history.
- Summarise back what you heard, check accuracy, and ask what you’ve missed.
Time yourself to whatever station length you’re practising, and say your reasoning out loud as you go so the transcript captures it. Keep that transcript, mark it against your structure, and track the phrasing you want to keep. Use fictional details only, never real patient information.
Step 4: make it mark you
When you say "end of consultation", paste a second prompt that asks ChatGPT to step out of character and mark you against the case file. You want it to audit completeness and accuracy, not vibes.
Step out of character. Compare my transcript against the case file above. List hidden facts I elicited and those I missed. List any red flags I failed to ask. Did I explore ICE. Was my closing summary accurate, and what key items were missing. Give two specific questions I should have asked next time. Keep feedback educational, no real patient advice. Assume this is a fictional case only.
A 2024 GPT-4 study found strong agreement with a human rater for structured completeness feedback, Cohen kappa 0.832. It fell below 0.6 in 8 of 45 categories, so some domains drift.
In a UK voice study, ChatGPT’s feedback came from a patient perspective rather than checking clinical accuracy. Helpful for rapport and clarity, weaker for clinical gaps.
So mark yourself against a grounded checklist as well. Build it from the negatives worth asking and from what examiners score in histories.
Where ChatGPT drifts off script
ChatGPT can role play a patient, but it will drift if you let it. Four patterns crop up. It fills gaps instead of saying I don't know. It volunteers hidden facts too early. It breaks character. And if you wrote the case, you already know the ending and will steer towards it.
It fills gaps. When you ask an unscripted question, it may invent plausible sounding answers. Tübingen 2024 logged 56.4% invented answers to unscripted questions, including socially desirable replies, leaving the patient role, or ignoring what the script already said.
It volunteers. Trying to help, it can hand over a hidden fact too early. That is why the prompt rule exists, and why you need to check the transcript after, to see whether anything leaked before you had actually asked for it.
It breaks character. In a UK voice study in 2025 it interrupted, ended scenarios early, or slipped into the doctor's role despite explicit instructions. This jars a novice and masks the timing and signposting you are practising.
- You wrote the case, so you know the answer. If you draft the scenario, you prime the outcome. You recognise your own cues and lead the witness. Ask it to write the case and it will often telegraph the diagnosis, or you guess it from what you requested.
What a purpose-built case adds
ChatGPT can role-play a patient and will keep up with a competent history. What it doesn’t give you out of the box is a fixed script with gated findings, labs and imaging you can actually order, and a scored commitment at the end, and that’s what a dedicated AI patient simulator is for.
First, the fixed script. In Diagnosica, Martin Rowe, 55, runs a small joinery firm. He opens with “I’ve suddenly gone short of breath with chest pain over the last two days,” and will volunteer sharp right-sided pain worse on breathing in plus sudden breathlessness.
The leg and the flight are gated. His left calf is swollen, tight, tender and warm for about a week only if you ask directly about the legs. The flight home from Australia five days ago, mostly seated, appears only when you ask about travel. Two streaky coughs of blood, also on a probe.
Miss those and the distractors feel plausible. Acute coronary syndrome or community-acquired pneumonia start to look tempting, with pneumothorax and pericarditis in the mix.
Second, you can order and see investigations. In that case you can request D-dimer, arterial blood gas, high-sensitivity troponin, a CT pulmonary angiogram, a 12-lead ECG, leg Doppler ultrasound and an echocardiogram.
Third, you have to stake your call. The skill being tested is committing to a ranked differential, then offering a plan, and getting a scorecard with competency scores and teaching points after the case.
It’s a video AI patient you talk to by voice or type. Every case is written and signed off by a doctor. There’s a free tier with no card.
"It doesn't give advice about real patients, and it's not a diagnostic system. Treat every output as educational."
A weekly routine that uses both
Use fictional cases only. Never enter real patient details.
- Voice session with ChatGPT on a case someone else wrote. Pick a published case with a mark scheme, run it as a timed interview, then mark yourself against the case file. If you want structure, use the full solo station method.
- Typed session with ChatGPT on a different case for pure questioning finesse. Work through Calgary Cambridge, ICE and SOCRATES, and force concise answers so you do the work. The free level costs nothing, so it is ideal for volume.
- One or two purpose-built cases (Diagnosica or another simulator) where you don’t know the diagnosis, so you have to find the hidden facts, order investigations and commit before you see the answer. Diagnosica has a free tier, and there are other free patient simulators.
- Optional third ChatGPT run, a timed redo of your hardest case. Keep the same mark scheme and compare against your first transcript to see drift or improvement.
- Review slot. Reread transcripts, highlight missed cues and waffle, then pick one habit to fix next week and write the exact prompt or checklist line you’ll use.
Run your first case tonight. No demo call, no card. Sign up, pick a case, and you are talking to a patient within a couple of minutes. Start free on Diagnosica
Educational use only, not medical advice. AI-generated; verify clinically against primary sources. Clinical review pending.
