Use AI for viva exam preparation as a rehearsal tool, not as an oracle. The best setup is simple: get it to generate realistic clinical stems, make you answer aloud under time pressure, then force it to critique your structure, prioritisation, and communication. If you only use it to produce polished model answers, you will sound smoother on screen and no better in the viva.
Why AI can genuinely improve viva performance
Examiners are usually listening for a few things at once: can you identify the priority, stay safe, explain your reasoning, and adapt when new information appears. AI is good at creating endless variation. That matters, because spoken exams are rarely lost on one fact; they are lost on muddled thinking, rambling, or missing the obvious first step.
That pattern spans current live assessments across specialties. The MRCGP Simulated Consultation Assessment uses twelve simulated consultations; MRCP(UK) Part 2 Clinical Examination (PACES23) mixes communication, examination, and consultation stations; MRCOG Part 3 uses a 14-task circuit with structured discussion and simulated patient or colleague tasks. Different formats, same pressure: think clearly, speak clearly, and stay safe.
AI helps most when you use it to rehearse that process. It helps least when you ask it to tell you the one perfect answer.
Best AI workflow for viva exam preparation
Start with stems, not model answers
Feed the tool topics, domains, and time limits. Ask for one case at a time, at the level of your exam, with no answer shown up front.
Useful prompt ingredients include:
- your stage: registrar, senior resident, fellow
- the exam style: consultation, structured discussion, bedside synthesis, ethics, communication
- the time pressure: 60 seconds, 90 seconds, 3 minutes
- the examiner style: neutral, interrupting, sceptical, safety-focused
- the output you want: stem only, follow-up questions, or post-answer critique
Act as a postgraduate viva examiner.
Give me one acute medicine viva stem for a 90-second answer.
Do not show the ideal answer.
After I reply, ask three follow-up questions that increase the pressure.
Then score me on:
- opening structure
- prioritisation
- safety
- clarity
- handling uncertainty
This is far better than asking for notes on sepsis or heart failure. You are training performance, not just recall.
Make the AI behave like an examiner
Your prompt should force challenge. Ask it to interrupt, add new observations, or question an unsafe plan.
For example, you can rehearse:
- a primary care phone consultation about a febrile 2-year-old, where safety-netting is the real test
- a physicianly consultation where you must explain next steps to a 67-year-old with new ascites
- an obstetric emergency discussion where you need to lead with immediate actions in postpartum haemorrhage
If the tool becomes too helpful, reset it. Tell it not to coach you mid-answer and not to reward vague language.
Debrief the transcript, not your memory
After a spoken answer, paste your transcript or rough notes and ask for a focused critique. The useful questions are usually small and brutal:
- Did I state the immediate priority in the first sentence?
- Where did I sound generic?
- What red flags or safety steps did I miss?
- Where did I over-investigate before stabilising the patient?
- Which phrase made me sound uncertain in a bad way rather than appropriately cautious?
Here is my answer transcript.
Identify:
- the first point where I should have been more decisive
- any missed immediate safety actions
- any unnecessary detail
- any jargon that would sound poor in a viva
Rewrite only my first 4 sentences so they sound clearer and safer.
That gives you something you can actually fix in the next repetition.
Turn weak spots into drills
The fastest gains come from recycling your errors. If you keep missing the same thing, make the tool generate five near-identical cases until the response becomes automatic.
Typical drill themes are:
- opening a management answer
- explaining risk to a patient or relative
- giving a focused differential after an examination finding
- handling the question what else are you worried about
- safety-netting and follow-up
- saying I need senior help early enough
Examiners notice pattern reliability. So should you.
Prompt patterns that actually work
A good AI mock viva is narrow. One task. One time limit. One debrief target.
For consultation-style stations, ask the tool to judge whether you opened with rapport, agenda, red flags, and a clear plan. For structured discussions, ask it to penalise answers that start with background instead of priorities. For bedside or clinical viva summaries, ask it to listen for three things: problem representation, likely diagnosis, and immediate next steps.
A simple answer scaffold works across specialties:
- Headline: what you think is going on, or what matters most right now
- Immediate priorities: what you would do first to keep the patient safe
- Key reasoning: the one or two facts driving your view
- Plan: investigations, treatment, escalation, and review
- Communication: what you would explain, check, or safety-net
If your answers routinely follow that arc, you already sound more exam-ready.
Use AI safely and ethically
Do not paste patient-identifiable material into a public AI tool. GMC confidentiality guidance says you should use the minimum necessary personal information and anonymise where practicable, and its reflective practice guidance reminds you that details such as age, job, ward, or condition can still identify someone when combined.
That means your prompts should use synthetic cases or heavily anonymised composites. Change the age, setting, timeline, and distinguishing details. Better still, ask the tool to create a fresh case that matches the learning point you need.
You also need human oversight. GMC guidance says doctors remain responsible for the decisions they take when using AI, and NHS England guidance on AI-enabled scribing stresses authorised use, review and approval of outputs, and appropriate human oversight rather than blind acceptance of the text produced.
For exam prep, that translates into one rule: never trust the tool on current exam policy, scoring rules, or factual detail unless you check the awarding body or guideline yourself. Use AI to rehearse; use official sources to verify.
Do not upload recalled live stations, copied candidate materials, or anything that breaches exam rules. For example, the current MRCP(UK) PACES23 candidate guide explicitly says case information must not be shared and scenarios must not be copied or removed from the centre.
Common mistakes
- Using AI to read notes at you instead of forcing you to answer aloud
- Asking for perfect model answers and then trying to memorise them
- Practising cases that are too long and too kind
- Accepting feedback like be more structured without asking where the structure broke
- Letting the tool reward unsafe overconfidence
- Ignoring tone, pacing, and signposting
- Rehearsing only your strong topics
- Pasting real patient details or workplace consultation audio into the prompt
A practical weekly practice workflow
If you are working full time, keep this lean. Short, repeated exposure beats occasional marathon sessions.
A workable week often looks like this:
- Two 20-minute AI sessions: one for rapid-fire stems, one for transcript debrief
- One 30- to 45-minute human mock: with a colleague, trainer, or study partner
- A running error log: three recurring weaknesses only, not twenty
- One rebuild session: rewrite and re-say the first minute of poor answers
A good debrief loop is simple:
- answer the case aloud
- self-score before reading any feedback
- compare your self-score with the AI critique
- rewrite the opening
- repeat the same theme with a slightly different stem
That last step matters. Improvement is not understanding the feedback. Improvement is sounding better on the next case.
Summary
- Use AI for viva exam preparation as a rehearsal engine, not a source of truth.
- Make it generate short stems, challenge questions, and targeted debriefs.
- Practise aloud under time pressure; silent review is low-yield.
- Turn repeated errors into narrow drills until your opening and prioritisation become automatic.
- Keep it safe: no identifiable patient data, no copied live exam material, and always verify official exam rules yourself.
References
- https://www.gmc-uk.org/professional-standards/learning-materials/artificial-intelligence-and-innovative-technologies
- https://www.gmc-uk.org/professional-standards/the-professional-standards/confidentiality
- https://www.gmc-uk.org/education/standards-guidance-and-curricula/guidance/reflective-practice/the-reflective-practitioner---supplementary-guidance-for-medical-students/anonymising-reflective-notes
- https://www.england.nhs.uk/long-read/guidance-on-the-use-of-ai-enabled-ambient-scribing-products-in-health-and-care-settings/
- https://www.rcgp.org.uk/mrcgp-exams/simulated-consultation-assessment/introduction
- https://www.rcgp.org.uk/mrcgp-exams/simulated-consultation-assessment/marking-and-results
- https://www.rcgp.org.uk/gp-training-and-exams/mrcgp-exam/mrcgp-exam-applications.aspx
- https://www.mrcpuk.org/sites/default/files/documents/Candidate%20guide%20notes.pdf
- https://www.rcog.org.uk/careers-and-training/exams/mrcog-our-specialty-training-exam/mrcog-part-3/mrcog-part-3-format/
- https://www.rcog.org.uk/careers-and-training/exams/mrcog-our-specialty-training-exam/mrcog-part-3/global-expansion-and-sustainability-project/
- https://www.gmc-uk.org/professional-standards/the-professional-standards/confidentiality/the-main-principles-of-this-guidance