How to Practice for OSCEs with AI Patient Simulations

TL;DR

An OSCE rewards one thing above all: reps with a patient who talks back, followed by specific feedback on what you missed. Standardized-patient sessions are the gold standard and you will never get enough of them, because they cost your school real money and real people. AI patient simulation fills the gap for the parts of the exam that happen in conversation: focused history taking, explaining a diagnosis, counseling, building a differential, presenting your findings. It does not fill the gap for physical-examination technique, and no AI tool does, because there is no body in the room. The evidence so far is encouraging but modest: a 2025 randomized trial in NEJM AI found students who trained with GPT-4-powered text-based simulated patients scored higher on a real OSCE than controls, by a small margin, and McMaster students described AI patients as a supplement to standardized patients, not a replacement. That is the right frame. Below: which stations AI simulation helps with, where the incumbents still win, and how Neural Consult’s Case Simulator fits into a four-week plan.

Why OSCE prep needs different tools than board prep

If you have spent the last year inside a question bank, the OSCE is a different sport. A multiple-choice item gives you the history; a station makes you extract it from a person who answers only what you ask, while a clock runs and an examiner ticks boxes. The skills being scored are observable behaviors: did you ask about onset, screen for red flags, explain the plan in plain language, wash your hands, summarize back. Knowing the answer is necessary and nowhere near sufficient.

Where those stations live has changed. In the United States there is no longer a national clinical-skills exam for physicians. The USMLE’s co-sponsors suspended Step 2 CS in May 2020 and announced on January 26, 2021 that they had no plans to bring it back, saying clinical reasoning and communication would still be assessed on the remaining Step exams and that they would work with schools on other ways to assess clinical skills. On the osteopathic side, COMLEX-USA Level 2-PE (administered 2004 to 2020) was suspended indefinitely in February 2021 and formally discontinued in June 2022; the NBOME says work has begun on a standardized-patient capstone, the C3DO, but it is not a licensing exam today. So for MD and DO students, OSCEs are now run entirely by your school, on its blueprint, scored with its checklists. The same is true for most PA, nursing, pharmacy, dental, and rehabilitation programs, which have used OSCE-format assessments for years.

Outside the US the OSCE is still a licensing hurdle. Canada’s NAC Examination, required of international medical graduates applying to Canadian residency, is a half-day OSCE of 12 stations (two unscored pilots), each 11 minutes with two minutes between, assessing history taking, physical examination, communication, diagnosis, investigations, data interpretation, and management. The UK’s PLAB 2 is 16 scenarios of eight minutes each. For those exams the blueprint is public; build your practice around it.

Whichever OSCE you face, the bottleneck is the same: many encounters with a patient who responds to you, and someone to tell you what you missed. Everything in this guide is about getting more of those two things.

What AI patient simulation can and cannot do

Start with the limits, because they decide how you spend your time. An AI patient is a language model playing a role from a case script. It can hold a coherent history, volunteer or withhold details depending on what you ask, answer an open question differently than a closed one, and react when you use jargon or skip the patient’s ideas, concerns, and expectations. Because it is software, it is available at 11 p.m., never tires of the same chest-pain case, and can score you against a checklist the moment the encounter ends. Those properties make it useful for history taking, counseling, explaining results, breaking bad news, building a differential out loud, and presenting a case.

What it cannot do is be examined. There is no abdomen to palpate, no reflex to elicit, no chest to auscultate, and your technique and sequencing are invisible to it. Some tools let you say “I would examine the cardiovascular system” and hand you findings, which keeps a case moving and lets you practice what you do with findings, but it is not physical-exam practice and nobody should sell it as such. An AI patient also cannot see you. Eye contact, posture, the pause before bad news, the tone that earns trust: standardized patients and faculty give feedback on all of that, and a text or voice model gives none. In the McMaster study below, students said exactly this: rapport was harder to build with the AI, and the SPs’ feedback on body language and tone was something the tool could not provide.

The randomized evidence

The strongest study to date is a randomized controlled trial published in NEJM AI in 2025 by Lavigne and colleagues. Roughly 250 medical students were randomized to standard OSCE preparation alone or standard preparation plus an “AI-Standardized Clinical Examination” platform offering 16 text-based, GPT-4-powered simulated patient encounters with automated scores and written feedback. On the real OSCE that followed, the AI group’s median station score was 11.4 out of 20 versus 10.7 for controls (p = 0.02), with no difference on an otoscopy station used as a control. The authors flag possible crossover, since control students could have used AI tools on their own. Read it plainly: a statistically significant but small gain, on conversational stations, from a text-only tool. A reason to use AI simulation, not a reason to skip your school’s SP sessions.

A 2026 randomized study in Frontiers in Medicine (80 internal medicine students at Peking University Third Hospital) found that four weeks of AI virtual-patient sessions added to standardized-patient consultations improved OSCE scores more than standardized patients alone (change of 11.9 versus 3.2 points, p < 0.01). The authors are candid that the combined group simply got more practice and feedback, so the AI’s own contribution cannot be isolated; the useful finding is that more reps plus more feedback worked, and AI is the cheap way to add reps.

The qualitative picture matches. A 2025 McMaster University study in International Medical Education had 24 students use an LLM virtual patient for history taking. Self-reported comfort rose after a single session (p < 0.001), most found the feedback comparable to that of standardized patients and preceptors, and students also reported lower fidelity than a real SP, occasional glitches, and two hallucinations, including one hallucinated piece of feedback. The authors’ conclusion is the one worth keeping: a supplement, not a replacement. None of these studies measured improvement beyond a few weeks, and none involved Neural Consult; they are evidence about the category.

The incumbents, honestly

Standardized patient sessions

Nothing else is a trained human who can be examined, who reacts to your face, and who debriefs you afterward. If your school offers optional SP practice or bookable skills-lab hours, take every slot and treat it as the exam. The only weakness is scarcity: a typical student gets a handful of formal encounters a year, and the feedback arrives once. Use AI simulation to walk into those scarce sessions having already made your beginner mistakes somewhere cheaper.

Peer practice and checklist banks

Practicing with a classmate, one of you holding a checklist, is free, builds the rhythm of a timed station, and is the only way to rehearse physical-exam technique on a real person outside the skills lab. Its weaknesses: your partner is as inexperienced as you, cannot play a patient and mark you at the same time, and tends to let you off the hook. Station banks help with the marking half. Geeky Medics, the UK-based clinical-skills publisher, lists more than 1,300 OSCE stations with examiner checklists, with a free subset. A checklist bank plus a strict friend is a serviceable physical-exam setup, and worth keeping even if you also use AI.

UK-built AI OSCE tools

The UK, where the OSCE is a finals and licensing format, has produced the most specialized AI OSCE products. Geeky Medics’ AI OSCE practice offers more than 800 AI virtual patients across history-taking, information-giving, and examination scenarios, by text or voice, with an AI examiner that marks you against the station checklist, and you can write or generate your own stations. Text consultations are free; voice sessions cost 4 of the plan’s AI credits (the Clinical Skills plan includes 100 a year or 25 a month, per Geeky Medics’ support pages), and that plan was listed at approximately £49.99 a year in a third-party comparison we could not confirm on Geeky Medics’ own pages. Quesmed’s OSCE-AI is built around the UK MLA’s Clinical and Professional Skills Assessment, offers text or voice AI patients marked line by line against a station-specific mark scheme, and costs £39.99 a year; its page quotes three different station counts (180+, 340+, 450+), so call it “a few hundred.” Both are strong for a UK exam. For a US or Canadian student their blueprints, drug names, and guidelines are a step removed, and neither covers the oral-presentation, nursing, or pharmacy case formats North American programs assess.

A general chatbot playing patient

You can paste “act as a 58-year-old man with chest pain” into ChatGPT and get a passable conversation, and plenty of students do. Two problems. The chatbot has no case script and no rubric, so it volunteers too much, drifts when you ask the same thing twice, and grades you generously if you ask it to grade you at all. And when you ask it afterward what the right workup was, you are back to the general-hallucination problem: a 2024 JMIR study found roughly 28.6 percent of references generated by one general chatbot were fabricated, and a 2025 study in Communications Medicine found general models repeated planted clinical errors up to 83 percent of the time. For free, unstructured reps it beats nothing; for scored practice tied to a checklist, it is not the tool. More in Neural Consult vs. ChatGPT.

Where Neural Consult fits

Neural Consult is a study platform for health-professions students, used by more than 75,000 of them, and its Case Simulator is the piece built for this job.

The Case Simulator as your between-sessions patient

Each case gives you a door note, then a patient you interview by voice or text, who answers like a patient rather than a textbook. During the encounter you can request examination maneuvers and receive findings, and order labs and imaging, so you practice what to do with information rather than just collecting it. At the end you commit to a differential and a management plan and receive a score with feedback on the interview itself (did you ask about onset, location, family history), on your diagnostic reasoning, on whether your orders were appropriate, and on your management. The scoring is specific enough to act on, and you can run the next attempt immediately.

Three details matter for OSCE prep in particular. The case bank is filtered by degree, organ system, and rotation, and the scenario types track how programs actually assess: MD-style OSCE encounters ending in a post-encounter quiz; SBAR cases for nursing students, who interview the patient, synthesize, and deliver a formal oral handoff to an attending; and pharmacotherapy consult cases for pharmacy students. Second, you can create unlimited custom cases from scratch or from uploaded files, so your school’s published blueprint or presentation list becomes your practice patients. Third, it is voice-enabled. Asking out loud under a clock, hearing an answer, and deciding what to ask next is the skill the exam tests, and typing is not a substitute.

What it is not: a physical-exam trainer. It cannot see or hear your manner, and it will not tell you that you stood too far from the bed. Use it for the conversation, the reasoning, and the presentation; use a human for the hands and the face.

Turning a missed case into something you remember

Many OSCE misses are not communication failures; they are a student who built rapport beautifully and forgot to ask about anticoagulants. The fix is to convert every case you fumble into material you will see again. The Question Generator builds unlimited board-style questions on any topic at a chosen difficulty, with explanations, so a missed pulmonary embolism case becomes a ten-question set on PE risk stratification that evening. The Flashcard Hub turns your lecture material into spaced-repetition decks, exportable to Anki, which is where checklists, red-flag lists, and counseling scripts belong if they are to survive until exam day. And the AI Lecture Notebook lets you upload your school’s own clinical-skills guides and mark schemes and ask questions against them, so the answer to “what does my program expect in a diabetes counseling station” comes from your program’s document, not a UK checklist or a model’s general impression.

Checking the medicine without the hallucination tax

After a case you will have questions: first-line therapy, the screening interval, what the guideline actually says. AI Medical Search returns cited answers, including PubMed sources, so you can check the claim rather than trust it. The platform is built on a grounded-retrieval architecture; that design is also why Neural Consult scored perfectly on the published USMLE Step 1, Step 2 CK, and Step 3 sample sets (119, 120, and 137 questions; details here), and why in a blinded study its generated questions were rated equal or superior to NBME items, with explanations scoring 4.4 versus 3.5 (p < 0.001). Two caveats: those are benchmarks of medical-knowledge accuracy, not of OSCE performance or of the simulator’s grading, and for nursing, PA, pharmacy, or rehabilitation students they are medicine benchmarks, not validation against your exam. More on what AI study tools can and cannot show in Do AI study tools actually work?

A four-week OSCE practice plan

Weeks 4 and 3 out: breadth and diagnosis. Get your school’s blueprint or, failing that, the presentation list from your clinical-skills course. Run one AI case per presentation, voice mode, timed to your station length, without looking anything up. Afterward read the feedback, write down every checklist item you missed, and put the medical gaps into questions or flashcards the same day. Five to seven encounters a week is enough to find your patterns; most students have two or three (skipping the social history, forgetting to summarize, never asking what the patient thinks is going on).

Week 2 out: depth and the hard conversations. Repeat your worst presentations and add the station types that are awkward to practice with a friend: breaking bad news, explaining a new diagnosis, medication counseling, an angry or anxious patient. Build custom cases from your own clerkship patients, anonymized. Nursing students, run SBAR cases until the handoff is automatic; pharmacy students, run the consult cases. Book every SP or skills-lab slot available for this week and next.

Week 1 out: integration with humans. Shift to peer practice with a checklist bank, concentrating on physical-exam stations, where AI cannot help and technique drifts fastest. Keep one or two AI cases a day to stay verbal and timed. Reread your missed-item list from week four; if the same three things are still on it, that is your checklist for the walk between stations.

Preparing for the NAC or PLAB 2 instead of a school OSCE? Lengthen everything, build custom cases from the published blueprint, and set the timer to 11 or 8 minutes from day one.

Recommended OSCE prep stack

The bottom line

The OSCE is a performance exam, and performance improves with reps and feedback. Your school’s standardized patients give you the best of both, in quantities too small to get good on. AI patient simulation is the cheapest way yet built to add conversational reps between those sessions, and the first randomized evidence says it helps, modestly, on the stations where conversation is the skill. It does nothing for your hands and cannot see your face, so keep the skills lab and a strict friend in the plan. Used that way, with misses turned into questions and cards the same day, four weeks is enough to walk into most stations having already made your mistakes somewhere that did not count. Education Technology Insights listed Neural Consult among its Top AI-Powered Medical Education Tools for 2026; what matters more is whether the next case you run fixes the last one.

Ready to run your first case? Start with Neural Consult, pick your degree and a presentation from your blueprint, turn on voice, and set a timer. For more on the thinking behind this approach, see Why every medical student should practice OSCE cases with AI and, for nursing students, How nursing students can practice patient interviews without clinical risk.