Can AI simulated patients prepare medical students for psychiatry?
Sixteen UK medical students and nine psychiatry teachers saw flexible, low-pressure practice in an AI simulated patient—but after only a ten-minute trial, they also flagged emotional realism, bias, safeguarding and over-reliance.
The short version
- What happened
- The qualitative study involved 16 medical students and nine psychiatry teachers in five focus groups after each participant used a research-team AI simulated patient for about ten minutes.
- Who it affects
- Medical students may benefit from a low-stakes place to rehearse difficult conversations, but patients and learners also need safeguards against stereotyped presentations, sensitive-data leakage and replacing supervised human contact with a cheaper simulation.
- What remains unproven
- Unreviewed, single-site qualitative study with 25 self-selected participants.
Keep in mind
Unreviewed, single-site qualitative study with 25 self-selected participants.
Participants used one research-team prototype for only about ten minutes.
Focus-group themes describe experiences and concerns; they are not prevalence estimates.
No comparator, blinded skill assessment, retention test, clinical behaviour or patient outcome was measured.
Student and teacher samples were demographically narrow and may not represent other institutions or regions.
Source analysis · 1 linked source. Primary analysed source date: 30 October 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
Participants saw a rehearsal tool, not a substitute for people
A qualitative UK study found cautious interest in AI simulated patients as a way for medical students to rehearse psychiatric interviews on their own time and without the pressure of a real encounter. Sixteen students from Kent and Medway Medical School and nine psychiatry clinical teachers joined five focus groups. Each session began with about ten minutes using a voice-based simulated patient built by the research team with GPT-4o Advanced Voice Mode, Streamlit and LangChain, followed by roughly an hour of discussion. Researchers recorded and transcribed the conversations and used reflexive thematic analysis with multicycle coding.
Students and teachers described repeatable, self-paced practice as potentially useful before real patient contact or for uncommon scenarios. They positioned it as an introductory or supplementary activity, not a replacement for actors, clinician-led teaching or patients. That distinction matters in psychiatry, where pauses, affect, trust and adaptation are part of the clinical skill. Participants reported that response delays, mechanical dialogue and limited emotional depth could disrupt rapport; one teacher observed that interacting with a robotic style could make the learner’s communication more mechanical too.[1]
Safeguarding and representation were curriculum questions, not side issues
The discussions surfaced risks that a technical demonstration cannot settle. Students asked how a practice system should handle disclosures involving self-harm, boundaries or safeguarding, what data it stores and who is accountable for its behaviour. Teachers raised representation and bias: a narrow set of generated patients could reinforce stereotypes instead of preparing students for the diversity of real services. Both groups wanted scenarios aligned with curriculum goals, learner level and institutional systems rather than a general-purpose chatbot placed alongside teaching without governance.
For educators, the useful finding is not that the tool improves learning—the study did not test that—but that perceived convenience and perceived realism may pull in opposite directions. A private, repeatable rehearsal could lower the barrier to practising a difficult conversation. Yet if the simulated patient supplies too much information, responds implausibly or rewards checklist-style questioning, it could rehearse the wrong habits. A responsible pilot would therefore define the skill being practised, use faculty-reviewed cases, set escalation and privacy rules, and keep human simulation and supervised patient contact in the pathway.[1]
The distinction that matters
What this study can—and cannot—support
The paper reports participant experience after a short prototype trial. It does not measure educational effectiveness.
Observed
25 people used one prototype for about ten minutes
Five recorded focus groups and thematic analysis
Perceived
Students and teachers discussed possible use
Flexible practice, alongside realism, bias and governance concerns
Not tested
No educational comparator or follow-up
No evidence of better skill, retention, examination or patient outcomes
The evidence is a map of concerns, not a learning outcome
The sample was small and self-selected: ten students attended one face-to-face group and six joined online; four were in year three and 12 in year four, with 14 women and two men. The teacher sample comprised six resident doctors and three consultants from South-East England, with seven women and two men. All participants encountered one research-team prototype for only about ten minutes. Their comments may identify design and governance questions, but they cannot establish how common each view is or whether longer use changes confidence, communication or diagnostic reasoning.
The medRxiv paper is unreviewed and comes from one well-resourced medical school; it excluded graduate-entry students and did not compare the AI exercise with actors, standardized patients or usual teaching. No examination score, observed interview rating, retention measure or patient outcome was collected. The authors reported no specific funding or conflicts of interest and said AI was used as the learning tool and for proofreading, not to analyse the data. What would change the assessment is a preregistered multi-school comparison with blinded skill ratings, follow-up, failure and safeguarding logs, and subgroup analysis across diverse learners and patient presentations.[1]