Practice with Simulated People

Pattern · All patterns
Evidence · 125 claims (58 for, 52 mixed, 15 against) · 60 studies (15 causal, 13 review, 12 quant-synthesis, 9 qualitative, 7 design, 4 theoretical), q1–q4 · 17 of 60 report an effect size · 120 claims rest on one study

Description and scope

Practice in which the learner performs a professional or interpersonal task with someone who plays a role: a standardized patient (a trained actor), a virtual patient, a simulated client, customer or colleague, avatar students in a classroom simulation, a pedagogical agent or virtual human, or a role-play partner played by a large language model. The decisions it covers are which performance to rehearse, how realistic and how responsive the simulated person is, how many attempts a learner gets, how feedback and debriefing follow each attempt, and how practice is carried over to real patients, clients, students or customers. It applies the debriefing principle and the deliberate practice principle, and draws on the transfer of learning principle.

It draws on three neighbouring literatures as well as the simulation literature: pedagogical agents and virtual humans (embodied characters whose appearance, voice and responsiveness shape how learners engage), teachable agents (rehearsing by teaching or coaching a simulated learner), and computer-supported collaborative learning (CSCL), where a scripted or AI partner takes a role in a dialogue and collaboration scripts structure who does what.

The study configurations below are evidence; the response-dependent policy and the doses are an untested design proposal. Use the page to gather better evidence about a learner group, not to certify that a learner "can communicate with patients" or "can manage a classroom" from in-simulation scores.

What the evidence does and does not test. The wiki's strongest evidence is two meta-analyses of simulation-based learning in higher education and the health professions, against no intervention or other instruction, and one meta-analysis of virtual patients against traditional teaching. They establish that simulation-based practice raises complex skills on average, and that the gain is smallest for communication skills. They do not isolate the simulated person from the rest of the simulation (mannequins and procedural simulators are pooled in), and they report few delayed or workplace outcomes. Two randomized trials test simulated people directly for communication: a virtual-human trainer for medical students against a computer-based module (blinded OSCE several days later) and a multi-agent LLM standardized patient against a structured non-LLM control. AI role-play partners are otherwise represented in the wiki by reviews, small design studies and uncontrolled pilots, and one review reports a trial in which AI-assisted simulation was inferior to human standardized patients for a hands-on skill. Teacher-education simulations appear in the wiki only as one small design-based study comparing a virtual simulation with authentic classroom video; the wiki records no claim about mixed-reality avatar classrooms of the TeachLivE or Mursion kind, nor about "approximations of practice" as a teacher-education design. Treat those configurations as plausible transports from the health-professions evidence, not as tested.

Case-based learning is the neighbouring pattern for analysing a written case; this pattern begins where the learner has to act in an exchange with a person who responds. AI tutoring covers an AI that teaches the learner; here the AI plays the person the learner must serve, persuade, assess or teach.

Inputs: establish the design brief

Field	Record or elicit
Learner and starting observation	Task-specific experience (novice, intermediate, advanced, or ?), prior practice with real or simulated people, and one short baseline encounter: what the learner says and does, what they miss, how they recover. Record confidence or anxiety separately from performance. Preserve unknowns.
The performance	The named interaction (history-taking, breaking bad news, a parent conference, a de-escalation, a sales or support call, a requirements interview, explaining a concept to a confused learner) and its kind: communication, diagnostic, procedural, collaboration, teaching. Communication and hands-on procedural skills have different evidence here.
The simulated person	Who or what plays the role (standardized-patient, scripted-virtual-person, llm-role-play, avatar-operated-by-a-person, teachable-agent), how they respond (fixed script, branching, generative), what they may disclose and under what conditions, and how their realism and consistency have been checked.
Context and activities	Setting and mode (online-self-paced, online-instructor-led, classroom, workplace-clinical, lab), time per encounter and in total, who gives feedback and debriefs (mentor, peer, the system, nobody), whether attempts can be repeated without penalty, and the technology (screen, voice, VR headset).
Learner-valued goal	Ask what the learner wants from the practice and why (for example, to stop freezing in a difficult conversation, to pass a practical exam, to be ready for a first placement). Record disagreement with the designer objective; negotiate rather than assume agreement.
Designer objective	The behaviours that count (a rubric with observable items), the minimum standard, and whether the target is the rehearsed scenario or a family of scenarios.
Outcome	A performance rated with a different person (another simulated person, a human standardized patient or a real encounter), rated by someone blind to condition where possible; knowledge, confidence and satisfaction recorded separately; immediate and delayed horizons. In-simulation scores are a process measure, not the outcome.

A request for "realistic AI role-play" does not specify these inputs. Ask for them before choosing a simulated person, a number of attempts or a predicted gain. Check first that the simulated person behaves as the role requires: that it discloses only what a real person in that situation would, that it does not drift out of role or correct the learner, and that a scored rubric item can actually be earned in it.

Sequence and conditional policy

1. Name the performance and the rubric with the learner. State the observable behaviours that count and show one strong and one weak example of them (a recorded encounter, a worked transcript, a demonstration). For novices, keep the example, and prefer a dialogue (someone asking and someone answering, such as a recorded encounter between a trainee and the simulated person) to a narrated monologue: in one experiment students who overheard a virtual tutee questioning a virtual tutor recalled more and asked more transfer questions than students who overheard a monologue [+W] (reported in a review, 48 college students, no effect size printed). For learners with experience, a short statement of criteria may be enough. About 10 minutes.
2. Baseline encounter. One short encounter with the simulated person, unscripted for the learner, recorded. Keep it short and treat it as a learning exposure as well as a measurement.
3. Feedback and a debrief after every attempt. Immediately after the encounter: specific feedback against the rubric items (what the learner did, which item it earned or missed, what to try), then a short debrief in which the learner says what happened, what they were trying to do at one or two decision points and what they will change. Use the debriefing principle for the debrief's structure. A score or a "wrong" flag without reasons is not this step.
4. Repeat the same scenario once. A second run-through of the same scenario, attempting the named change, followed by feedback again. Record the change between attempts.
5. Vary and harden across sessions. In later sessions use new scenarios from the same family, with different people, presentations and complications (an emotional reaction, a withheld fact, a misconception, an interruption, the behaviour a real client of that kind would show), and fade the examples and prompts. Move a learner on when they meet the standard on a new scenario, not on a repeated one.
6. Probe transfer with a different person. At the end of the unit, and again after a delay where it matters, rate a performance with a different simulated person or a human standardized patient, ideally by a rater blind to how the learner practised; where real encounters follow, collect a rating from them.

Dose and progression (proposal, untested as a whole). Encounters of roughly 10–20 minutes, each followed by 5–15 minutes of feedback and debrief; two or three encounters per session; three to six sessions spread over the unit for one family of interactions. Early sessions: one scenario at a time, examples and rubric prompts visible, the same scenario repeated once. Middle sessions: new scenarios, fewer prompts, the learner self-assesses against the rubric before seeing feedback. Late sessions: unannounced complications, no prompts, a different simulated person, and the transfer probe. In the meta-analysis below, simulations lasting several hours to a day showed larger effects than ones under an hour, and duration beyond a month was not established; the 11-minute median session observed in one LLM role-play deployment for novice teachers is a description of use, not a recommended dose.

Observation	Discriminating probe	Proposed action, with its uncertainty
Learner knows what to say when asked, but does not do it in the encounter	Ask them to name the rubric behaviour for the moment just passed; rerun the same moment.	If they can name it but not do it, the gap is performance: give more repeated attempts with feedback on that item. If they cannot name it, give the example again first. Neither branch is a validated diagnosis.
Scores improve on the repeated scenario but not on a new one	A new scenario of the same family with a different simulated person.	Expected rather than a failure: in one trial scores improved sharply between run-throughs of the same scenario, and repetition in language tasks improved the repeated task without transfer to a new one. Add varied scenarios before moving on.
Learner is anxious or freezes	Ask what they fear in the encounter; compare a low-stakes private attempt with an observed one.	Offer private, repeatable attempts without grade penalty first, then observed ones. Reviews report AI practice as lowering anxiety, but one trial found higher state anxiety with AI simulation than with a standardized patient, so check rather than assume.
The simulated person answers everything, volunteers information, corrects the learner, gives way as soon as the learner pushes, or drifts from the role	Rerun with a probing question the real person would deflect, and with a request the real person would refuse; compare two runs of one persona.	Fix or replace the simulated person before scoring learners on it. An adaptive patient that disclosed more as the trainee explored behaved more realistically than a prompt-only one; LLMs prompted to play novices may not hold a misconception; and an argument holds that chat-tuned models are inclined to agree and may not push back as a difficult client would (untested in the wiki).
The simulated person stays in role but gets facts of the domain wrong or leaves content out	Score its turns against the content the scenario must convey (the facts a patient or client would know, the clinical details), separately from how well it plays the role.	Check content accuracy as its own step. In an evaluation of eight AI teaching agents in medical education, agents differed most on knowledge coverage and accuracy and least on role-play quality, and an agent revised to sound more empathic covered the least content; a persona that sounds right is not evidence that the content is.
Learner interrogates the simulated person like a database: closed, jargon-laden questions, no rapport, no explanation back to the person	Count open against closed questions in the transcript; ask the learner whether they were talking to a patient or querying a case.	Make the rubric reward open questions, rapport and explaining a plan to the person, have the simulated person answer closed questions narrowly, and add a closing phase in which the learner tells the person the outcome. In one mixed-methods study of a generative-AI patient in a problem-based tutorial, students treated the avatar as a question base and asked mostly closed questions [-W] (two tutorial groups, descriptive).
Good communication items, weak diagnostic or technical items	Score the domains separately.	Add targeted practice on the weak domain; the multi-agent trial raised communication scores without changing diagnostic accuracy, and simulation effects were largest for technical and smallest for communication skills.
High satisfaction and confidence, no change in rated performance	Blinded rating of a performance with a different person.	Do not read satisfaction or confidence as learning. Revise the feedback or the scenario family, and keep the rating as the outcome.
Learner fails a hands-on component practised only with an AI or screen-based person	Observe the component with a human standardized patient or a physical model.	Move the hands-on component to a human or physical simulation. One trial in a review found AI-assisted simulation inferior to a standardized patient for a breast self-examination skill.

Choosing configurations from evidence

• Whether to expect a gain from simulation-based practice at all. Simulation-based learning has a large positive overall effect on complex skills in higher education [+M]: a meta-analysis of 145 studies (10,532 participants) found g = 0.85, with no significant difference between waiting controls (g = 1.02), differently instructed controls (g = 0.82) and pre-post baselines (g = 0.88). Simulation-based education improves outcomes [+M]: across 609 health-professions studies (35,226 trainees, 137 randomized), technology-enhanced simulation against no intervention gave pooled effects of about 1.1–1.2 for knowledge and skills, about 0.8 for behaviours, and 0.50 for direct effects on patients (k = 32); it found no consistent interactions with design features such as feedback or repetitive practice, and heterogeneity was large. Neither isolates a simulated person from mannequins and procedural trainers, and the no-intervention comparison overstates what to expect against other good instruction.
• Which performance. Simulation effects vary by complex skill [~M]: in the same 145-study meta-analysis, communication (g = 0.44) and collaboration (g = 0.50) gained least, technical performance (g = 1.06) most, with diagnostic skills (g = 0.82) and problem solving (g = 0.88) between. Practice with simulated people mostly targets the skills with the smallest average gains, so plan more attempts and better feedback for them rather than fewer.
• Simulated people for communication: the two direct trials. In a blinded multisite randomized trial, second-year medical students who practised with MPathic-VR, a virtual-human trainer for intercultural and interprofessional communication, scored higher than students who completed a computer-based learning module on an advanced communication OSCE several days later, rated by standardized-patient instructors blind to condition (composite OSCE scores [+M], η² = 0.0145, means .806 vs .752), with the clearest advantage on nonverbal communication [+M] (η² = 0.032). Within the trainer, scores improved sharply from the first to the second run-through [+M] (η² = 0.45 and 0.36 for the two scenarios), and students reported more positive attitudes [+M]; in their reflective essays they valued the immediate feedback by video review and the practice of nonverbal skills and emotionally charged encounters, while the comparison group described the module as facts-based and lacking interactivity [+M]. So the within-simulation gain was large and the transfer gain on a different, human-rated encounter was small; plan for that gap. In a randomized study of medical students (47 vs 48), a multi-agent LLM standardized patient with scaffolding raised final OSCE-aligned examination scores [+M] (71.8% vs 55.6%, Hedges' g about 0.8) against a structured non-LLM control, most in communication [+M] (3.53 vs 2.64 on a 1–5 scale), but did not change diagnostic accuracy or worksheet completion [~M] (86% vs 84%; equivalence not tested). It is a preprint with one site.
• Virtual patients: replacing or adding to teaching. A meta-analysis of randomized trials found virtual patients no different from traditional education for knowledge [~M] (SMD 0.11, 95% CI −0.17 to 0.39, n = 927) but better for skills [+M] (SMD 0.90, 0.49 to 1.32, n = 897; 9 of 14 studies favoured them). Added to traditional teaching, they improved knowledge [+M] (SMD 0.73, n = 439, 5 studies) while the skills effect was positive but its interval crossed zero [~M] (SMD 0.60, −0.07 to 1.27, n = 247). Heterogeneity was high throughout (I² 74–88%). Seven of the 51 trials were run in low- and middle-income countries, which the authors read as evidence of applicability beyond high-income settings [+W] (a distribution of trials, not a subgroup test). The authors' exploratory reading is that the advantage is larger over passive teaching (lectures, reading) than over active learning [~M], and the few comparisons with other digital education found no significant differences [~M]. Use a simulated person to practise a skill, and do not expect it to beat an active alternative already in place.
• AI role-play partners against human standardized patients. A mixed-methods review of AI-powered simulation in nursing found its three trials support short-term gains in communication confidence and structured knowledge with limited or inferior effects for tactile psychomotor skills [~M]; one trial there found AI-assisted simulation inferior to standardized patient simulation for breast self-examination skills [-M], and AI simulation produced higher state anxiety than a standardized patient in one trial while qualitative studies reported less [~M]. A scoping review of generative AI in medical education reported short-term objective gains in some controlled virtual-patient studies, with positive findings not universal [~M]: some implementations measured only confidence or usability, and human or conventional cases were sometimes rated more educational or authentic. The same nursing review's subgroup synthesis reports that generative-AI and LLM partners carried the strongest evidence for cognitive and communication outcomes, AI-driven mannequins mixed evidence for psychomotor skills, and stand-alone chatbots weak independent effects [~M] (narrative synthesis of 19 studies, no pooled effect), and it describes the typical AI outcome: a pre/post rise in clinical confidence [+W] (1.06 points on a 5-point scale, no comparison group), which is confidence and not performance. A review of live-action simulations reports that face-to-face formats add embodiment and social immersion, eliciting more collaborative talk, while virtual formats make systematic trials and resets easier [~M]. Keep human or physical simulation for hands-on and embodied components, use AI partners for conversational ones and for the repeatable attempts a live format makes costly, and do not offer a chatbot alone as the practice.
• Repetition and deliberate practice. Simulation-based education with deliberate practice improves clinical outcomes [+M]: across 14 studies, simulation with repeated, goal-directed practice, mastery standards and immediate feedback outperformed traditional clinical education (d = 0.71, 95% CI 0.65–0.76), mostly on procedural skills. Longer simulations were associated with larger effects [+M] (under an hour g = 0.65; several hours to a day g = 0.81; up to a month g = 0.80; over a month, 4 studies, inconclusive). The task-repetition evidence is a warning: rehearsal improved fluency and complexity on the repeated task but did not transfer to a new task of the same type [~M] (second-language speaking tasks). In a self-service study of an AI communication trainer (59 participants, two encounters with different AI patients in counterbalanced order), scores were higher in the second encounter [~M] (0.403 to 0.471 on an LLM-judged rubric; one rubric dimension did not change), which suggests some carry-over to a new patient, but there was no control group and the scores were machine-rated. Hence step 4 repeats once and step 5 varies.
• Feedback after each attempt. Feedback improves learning by a medium amount on average, but more than a third of feedback interventions reduce performance [~S]: high-information feedback (d = 0.99) outperformed corrective (d = 0.46) and reinforcement-only (d = 0.24) feedback. In a qualitative pilot with ten graduate speech-language pathology students, simulation feedback that marked answers wrong without explanation was judged insufficient [-M], while the same students valued the debrief with the clinical instructor [+W] and unlimited attempts without grade penalty [+W]. Among virtual-patient design variants, which mostly showed no differences, empathic feedback improved communication skills over no feedback [~M]. Medical students using an LLM standardized patient valued repeatable practice and recorded review but questioned missing actions and feedback criteria [~M]; their peers described human SP practice as valuable for embodied cues but episodic and unevenly connected to feedback [~M]. Interviews and co-design workshops with 12 clinical-year medical students produced four design themes for AI standardized-patient practice [+W]: treat human SPs as the reference, extend practice beyond the dialogue (examination, ordering, documenting), connect the activities of an encounter, and support review after the session. In a two-psychologist expert evaluation of a psychotherapy virtual patient, turn-by-turn feedback with a retry option increased awareness of intervention choices [+W], a finer-grained variant of feedback after each attempt; the wiki records no comparison of the two. Where an instructor plays the simulated person, a design ethnography of maritime simulator training found corrections given first in role and then, when students showed no sign of understanding, restated out of role by the instructor, and concluded that recurring movements between the in-role and the out-of-role context are what let simulation teach [~W] (qualitative, one course): so a correction given only in character may not register, and the person giving feedback should step out of the role to say it plainly.
• Scaffolding by prior knowledge. In the 145-study meta-analysis, adding a single type of scaffolding did not raise effects, and prompts alone were associated with lower effects (g = 0.65 with prompts vs 0.92 without), but some combinations did better, notably examples with prompts in four medical-education studies (g = 1.60) (scaffolding added value partly supported [~M]). Prior knowledge moderated which scaffold helped [~M]: reflection phases did more for learners at higher education levels (g = 1.10), examples for those at lower levels (g = 0.88, against 0.52 for reflection). Hence the example-first start for novices and the shift to self-assessment and reflection later.
• Authenticity and fidelity. High-authenticity simulations showed larger effects than low-authenticity ones [+M] (g = 0.86, N = 76 vs g = 0.58, N = 26), and document-based simulations were the least effective type [~M] (g = 0.31, N = 15). More technology support was associated with larger effects [~M] (simulators in medicine g = 1.07; VR g = 0.85; screen-based g = 0.68; none g = 0.74), an association across study categories, not a within-study contrast. Authenticity here is how fully the situation is represented, not visual polish: a review of serious games reports that higher perceptual fidelity does not reliably improve learning [~M] and can add load and cost. In an embodied clinical-communication study (59 participants), a conversational model that participants judged more human-like produced the same measured performance [~M] (0.437 vs 0.438). Responsiveness is the property with direct evidence: an adaptive virtual patient whose disclosure climbed across a session [+M] responded to the trainee's exploration but not measurably to their empathy [~M], where a prompt-only patient with the same model stayed flat (20 clinicians and trainees, within-subjects). Learners notice missing behaviour: most of ten speech-language pathology students felt simulated child-assessment cases lacked authenticity because they included too few behavioural complications [-W], which is the reason step 5 adds the behaviour a real client would show. A conceptual framework for agents in extended reality predicts that learners treat immersive agents as real people [+W] and that behavioural realism should predict social presence [+W]; these are predictions, not tests, but they point the same way as the responsiveness evidence: how the simulated person behaves matters more than how it looks. Behaviour and appearance should also match: in a study of an AI writing partner (48 participants, three variants), the variant whose timing was humanlike but whose appearance was not was rated least positively on partnership and agent perception [~W] (exploratory tests, perceptions only), with the fully humanlike and the fully tool-like variants rated alike.
• Checking the simulated person's content as well as its role. An evaluation that scored 167 dialogues from eight AI teaching agents used in medical education found the agents differed most on knowledge coverage and medical accuracy and least on role-play quality, with difficulty calibration weak across nearly all of them [~M], and an agent whose prompt was revised to be more empathic covered the least content [~M] (scored by a language model, one agent's case). An argument from how chat models are trained holds that they tend towards agreement and may not push back [-W], which matters where the role is a sceptical client or a resistant patient. So the pre-use check has three parts: stays in role and discloses as a real person would, holds the facts of the scenario, and resists where the real person would.
• Pedagogical agents and virtual humans: appearance and voice. A retrospective review reports a meta-analysis of 43 studies in which pedagogical agents enhanced learning compared with environments without agents [+M] (no effect size printed). Reviews attribute the effect mainly to the agent's voice rather than its visual presence [~M], report animated personas outperforming static ones, with voice sometimes contributing more than the animation [+M], and report that facial expressions helped attitudinal learning but hurt procedural learning [~M]. Agents as coaches and tutors are reported to be most useful for novices, while virtual role-players benefit learners across expertise levels [~M], an interpretive summary rather than a single test. Where the simulated person is a fellow learner or companion rather than a client, its competence matters: in a 2×2 experiment with 72 undergraduates, working with a high-competency learning companion improved application [+M] (d = 0.46) while working with a low-competency companion raised self-efficacy [~M] (d = 0.49, a one-item measure), so the two goals pull in opposite directions. Ratings of a persona can carry learners' stereotypes about it rather than its quality: in studies where agents differed only in presented gender, with identical scripts, learners rated the male-presented agent more positively and recalled more after working with it [~W] (reported in a review, no effect sizes), which the authors read as gender stereotyping by learners, so judge a simulated person by what learners do with it, not by how they rate it. In a workshop study, 31 teachers' ratings of an AI group partner's proficiency, etiquette and personality did not differ detectably across four voice-accent conditions [~W] (small samples per condition, so an absence of a detected difference). Give the simulated person a voice before a face, and add facial expression where the target is attitudinal or emotional rather than procedural.
• Teaching a simulated learner (teachable agents and simulated students). Learning by teaching has the firmest evidence of the configurations here, though with human tutees: preparing to teach and teaching improved the tutor's own learning [+M] (28 studies; g = 0.56 for teaching after preparation, 0.35 for preparation alone, larger when the teaching was interactive). With simulated learners, adding a teachable agent to a fourth-grade science curriculum improved learning gains [+M] (seven classrooms, quasi-experimental, d = 1.1 and 0.93), and university students who taught a ChatGPT agent gained more programming knowledge than students learning from videos [+M] (20 vs 21; the printed η² = 0.74 overstates the effect) but not more code correctness [~M]. For teacher candidates rehearsing with simulated students, teachers in an AI-guided professional development programme said simulated student errors moved them from solving tasks to reasoning about students' thinking [+W] (qualitative). Whether an LLM can play a realistic novice is itself unsettled: an argument holds that capable LLMs cannot genuinely unknow expert knowledge [-W], and in one test multi-turn interaction did not make LLM simulators hold a misconception faithfully and destabilised weaker ones [-M], while simulators trained on real student data outperformed prompted LLMs on behavioural fidelity [+M] in chess, second-language writing and mathematics. The same trained simulators responded better to a tutor's guidance than every baseline, most clearly when the guidance was a question rather than the answer [+M], and in one held-out chess case a small trained simulator followed Socratic guidance that a prompted frontier model did not [+W]. On a mathematics tutoring dataset, simulators given a profile built from each student's history matched students' turns better than prompting or plain fine-tuning [+M] (small gains), and a model made to "unlearn" content and then taught by a coach agent recovered more, and more progressively, than one relearning from static material [+W]. Training is not always better: for a writing prompt with few training essays and long responses, prompting outperformed the trained simulators [-M]. Simulated learners used to test a system can also mislead: a tutor that looked adaptive in simulation was judged adaptive by only three of eight real learners when they expressed uncertainty [-M]. A simulated persona's self-reported traits stayed stable across a long conversation [+M], but model family and prompt format changed how strongly the persona was expressed [~M]. Validate a simulated learner against real learners' responses before training teachers on it.
• Simulated partners in collaboration scripts. The wiki's evidence on CSCL scripts comes from two meta-analyses of human groups by one research group, not of AI partners. Against unscripted collaboration, scripts raised domain learning a little [+M] (d = 0.20; g = 0.24 in the update) and collaboration skills more [+M] (d = 0.95, 0.65 without two outliers), which is the case for scripting the roles a simulated partner plays. In the moderator analyses, the script level [~M], whether the script prompts transactive talk [~M] and added content support [~M] did not significantly moderate effects (with post hoc power of 8–16%, so these are absences of detected differences), while subject discipline moderated effects on collaboration skills [~M]. Where an AI partner takes a scripted role, the evidence does not say which role or script works best.
• Teacher education: simulation against authentic video. In a design-based study with preschool teacher candidates (eight groups), authentic classroom video produced higher observation-assignment scores than a virtual simulation [-M] (85.0 vs 81.75, d about 1.2), with the advantage on analysis and support dimensions [-M] and no difference on objective description [~M], while the virtual simulation prompted more self-reported purposeful reflection [+M] (d = 0.51). The order was fixed (simulation first), so the comparison is confounded with sequence. For teacher candidates learning to read classrooms, pair simulation with real footage rather than replacing it.
• Transfer and horizon. No study recorded here follows practice with simulated people to a delayed workplace outcome beyond the 32 patient-outcome studies pooled in the health-professions meta-analysis. Reviews report that simulation games support near transfer, with far and long-term transfer less robust [~M], that simulation advantages over physical labs on immediate tests did not persist to later examinations [~M] (physics, not people), and that simulation games helped as a supplement (d = .51) but not as standalone instruction (d = −.12) [~M]. Nurses interviewed after specialty training said that the closer training was to the clinical scenarios they would meet, the more smoothly it carried into practice [+M] (qualitative), students in a simulated clinical course recommended a simulation case aligned with each course's content [+W], and a review of situated learning argues that immersive technology should reinforce, not substitute for, the original educational action [~W]. Plan the transfer probe and a delayed probe, build scenarios from the situations learners will actually meet, and embed the simulation in a course rather than offering it alone.

Do not rank these results against one another: comparators (no intervention, lectures, computer modules, structured non-LLM practice, human standardized patients, authentic video), populations and outcomes (OSCE ratings, in-simulation scores, knowledge tests, confidence) differ. A precise abstention names the missing comparison, outcome or observation and the next way to resolve it.

Fitting the design to a situation

Ask for four facts when the brief omits them: what exactly is the performance (communication, diagnostic, hands-on, teaching), and how will it be rated with a different person?, who or what plays the simulated person, and has it been checked to stay in role?, who gives feedback and leads the debrief, and how soon after each attempt?, and how much experience do learners bring to this kind of encounter?

If the brief says…	Then change…	Basis
Learners are novices in this kind of encounter (novice)	Start with a modelled example, shown as a dialogue rather than a narration, and rubric prompts; move to self-assessment and reflection phases only after they meet the standard on a new scenario	Prior knowledge moderated which scaffold helped [~M]; agents as coaches most useful for novices [~M]; overheard dialogue beat monologue [+W] (one experiment, reported in a review)
Learners are experienced practitioners (advanced)	Drop the worked example, use harder scenarios with complications, and put the time into reflection after each attempt	Reflection phases did more at higher education levels [~M]
Learners are school-age, rehearsing by teaching	Use a teachable agent inside the regular curriculum, not as a separate activity; give teachers planning time, a curriculum map and a rehearsal of their own before the first lesson	Teachable agent improved fourth-graders' gains [+M]; carried from one quasi-experiment; prep time, curriculum alignment and teacher confidence are reported barriers to K-12 simulations [-W] (teachers' reports, via a review)
School-age learners would practise social or emotional situations with an AI character	Keep it a supervised, bounded rehearsal with an adult debrief; set age limits locally, and check the tool's reliability before pupils use it; have the character disagree, refuse and disappoint as a real peer would rather than please, and move the rehearsal on to people	Teachers describe AI companions as a low-stakes rehearsal space but tie acceptability to age [~W]; primary teachers worry about AI reliability, age appropriateness and over-reliance [-W]; teachers' most common concern is that agreeable companions remove the friction through which adolescents learn to sustain relationships [-W]; teachers' views, not outcomes
The goal is confidence in the task as well as skill, and the simulated person is a fellow learner or companion	Pair a peer-level companion for early confidence-building sessions with a more competent one, or a stronger model, for sessions aimed at application; measure self-efficacy and performance separately	High-competency companion improved application [+M]; low-competency companion raised self-efficacy [~M]; one experiment with undergraduates, the sequence untested
The goal is communication or interpersonal skill	Plan more attempts and richer feedback than for technical skills; rate the outcome with a human, blind to condition	Communication gained least from simulation [~M]; MPathic-VR OSCE transfer [+M]
The goal includes a hands-on or physical-examination skill	Practise that component with a human standardized patient or a physical model, not an AI or screen-based person	AI simulation inferior to SP for breast self-examination [-M]; limited effects for tactile skills [~M]; AI mannequins mixed for psychomotor skills, LLM partners strongest for communication [~M]; face-to-face simulation adds embodiment [~M]
The goal is diagnostic or clinical reasoning as well as conversation	Score reasoning separately and add reasoning-specific practice; do not expect the conversational practice to carry it	Multi-agent SP did not change diagnostic accuracy [~M]
Online, self-paced, no live facilitator (online-self-paced)	Build specific, rubric-referenced feedback and a written debrief prompt into every attempt; allow unlimited low-stakes repeats; add a mentor review of a recorded encounter at the end of each scenario family	High-information feedback outperforms corrective feedback [~S]; feedback without reasons judged insufficient [-M]; learners value repeatable practice and recorded review [~M]; completers rated a self-service AI trainer engaging and easy to use [+W] (feasibility, not learning)
Mentors or faculty have little time	Use an automated assessor for formative scoring of every attempt, and spend mentor time on the debrief of selected recordings and on the transfer rating	An LLM assessor agreed with human raters about as well as individual human raters [+M] (one validation, one rubric); the division of mentor time is untested
Time is short (one session)	Run at least one baseline encounter, feedback and debrief, and a repeat; tell learners the gain on the repeat will be larger than on a new scenario	Large first-to-second run-through improvement [+M]; repetition did not transfer to a new task [~M]
Time allows several sessions	Spread scenarios over sessions and vary them; longer total practice is associated with larger effects	Longer simulations associated with larger effects [+M]
A traditional, lecture-based component is being replaced	Replace the passive component with simulated-person practice; expect less from replacing an active one	Larger advantage over passive than active teaching [~M]; VPs better for skills, not knowledge [+M]
Stakes are high (certification, licensure, a practical exam)	Use the simulated person for practice; keep the decision on a rated performance with a human standardized patient or examiner, and on the transfer probe	untested proposal; human or conventional cases were sometimes rated more authentic [~M]
Learners report anxiety about practising in front of others	Begin with private AI attempts, then observed ones; measure anxiety rather than assume it falls	AI practice reported as lowering foreign language anxiety [+M]; nursing learners valued safe, repeatable, nonjudgmental AI practice [+W] (qualitative, 11 studies); higher state anxiety with AI than SP in one trial [~M]
The plan uses VR headsets	Keep sessions short, give a screen alternative and check for symptoms; use VR to add to live or real practice, not to replace it	Cybersickness reported in about 42% of VR users [-M]; VR should reinforce, not substitute [~W] (a review's interpretation)
The simulated person is a prompted general LLM	Test persona consistency and disclosure rules before use, compare two runs of one persona, score its content accuracy separately from its role-play, test whether it resists where the real person would, and prefer a responsive design that discloses as the learner explores	Adaptive patient's disclosure climbed while a prompt-only one stayed flat [+M]; model and prompt moderate persona expression [~M]; agents differed most on knowledge, least on role-play [~M]; sycophancy may limit push-back [-W] (argument)
The budget is small or the cohort large	Use an LLM role-play partner for the repeated attempts and spend the money saved on checking the simulated person and on mentor debriefs and transfer ratings	LLM classroom role-play cost $0.05–$0.10 a session [+W] (one deployment, 150 users); costs of checking and debriefing not recorded in the wiki
The goal is speaking in a second or foreign language	Use an AI conversation partner in simulated everyday or workplace scenarios with corrective feedback after each turn or exchange; vary the scenarios rather than repeating one, and rate speaking with a person at the end	AI corrective feedback and repeated practice reported to improve speaking [+W]; AI simulates authentic interaction scenarios with tailored scaffolding [+W] (both a systematic review's thematic coding, no effect sizes); repetition did not transfer to a new task [~M]
The simulated person is used inside a problem-based or case tutorial	Allow more time for the history (about ten minutes more in one study), add a phase in which learners explain the diagnosis and plan to the simulated person, and prompt open questions; do not expect better recall of the case from the simulated person alone	GenAI tutorials ran longer, and only those groups told the patient the diagnosis and plan [~W]; students treated the avatar as a question base [-W]; recall did not differ by condition [~W]; learners rated the tutorial higher [+W] (one mixed-methods study, four tutorial groups)
Health professions (history-taking, breaking bad news, interprofessional talk)	Use the virtual-human or LLM standardized patient for conversational practice with an OSCE-style rubric, a repeat run-through and a blinded OSCE at the end	MPathic-VR OSCE advantage [+M]; multi-agent SP raised OSCE-aligned scores [+M]
Teacher education (managing a class, responding to student thinking)	Pair simulated students with authentic classroom video and real placements; use the simulation for rehearsing responses and reflection, and the video for learning to read classrooms; validate simulated students against real student responses	Authentic video beat virtual simulation on analysis [-M]; simulation prompted more reflection [+M]; LLMs may not unknow expert knowledge [-W]; novice CS teachers who used an LLM classroom simulation viewed it positively [+W] (optional feedback, no outcome measure); mixed-reality avatar classrooms are untested in the wiki
Business (sales, negotiation, customer, management or coaching conversations)	Use the health-professions sequence with a business rubric; rate the outcome in a role-play with a different person or a real call, and treat the expected gain as the communication-skill estimate	untested in business settings in the wiki; transported from communication gained least [~M]
Technology (requirements interviews, user support, explaining code, code review)	For explaining and reviewing code, have learners teach a simulated learner and then check correctness separately; for client interviews, use a simulated client with rules for what it discloses	Teaching a ChatGPT agent improved programming knowledge [+M] but not correctness [~M]; simulated clients untested in technology settings in the wiki
An AI partner takes a role in a group discussion or collaboration script	Give the AI a defined role (clarifying the stage goal, keeping the group on topic, asking the questions people hesitate to ask), tell the group what that role is, check that it tracks the group's decisions accurately, and debrief the group, not only individuals	LLM participants clarified stage goals and prompted reflection [+W] (focus groups); an AI partner whose role was ambiguous and which tracked a group's decisions inaccurately lost the group's trust [~W] (one group's case study, school students); scripts improved collaboration skills [+M] (human groups); script features did not moderate effects [~M] (human groups)

The evidence by question

• Does simulation-based practice raise complex skills, and which ones? Simulation-based learning has a large positive overall effect on complex skills [+M]; Simulation effects vary by complex skill [~M]; Simulation-based education improves outcomes [+M]
• Does practice with a simulated person transfer to a human-rated encounter? MPathic-VR OSCE composite [+M]; Multi-agent SP and OSCE-aligned scores [+M]
• Should a virtual patient replace or supplement teaching? Knowledge: no difference [~M]; Skills: advantage [+M]; Blended: knowledge advantage [+M]
• Can an AI partner stand in for a human standardized patient? Short-term gains, inferior for tactile skills [~M]; Inferior for breast self-examination [-M]
• Which scaffolds, for whom? Scaffolding added value partly supported [~M]; Prior knowledge moderates scaffolding [~M]
• How realistic must the simulated person be? High authenticity [+M]; Fidelity fallacy [~M]; Adaptive disclosure [+M]
• Will the simulated person behave and inform as the role requires? Agents differ on knowledge, not role-play [~M]; Misconception faithfulness [-M]; Trained simulators and guidance [+M]; Simulation overestimated adaptiveness [-M]
• Does teaching a simulated learner help the one who teaches? Learning by teaching [+M]; Teachable agent in fourth grade [+M]

Further evidence, not yet read against this model

Read against the model on 2026-10-19; these bear on the page but do not change a configuration, a situation row or the default. - AI-driven case simulations in ambulatory settings support diagnostic reasoning in nursing education [+W]: a review's narrative statement with no effect size, about AI case simulations with diagnostic suggestions rather than a simulated person; it does not outweigh the trial in which a simulated patient left diagnostic accuracy unchanged, so the diagnostic row stands. - Students report more assessment confidence after a simulated clinical course [+W]: confidence from the same ten-student pilot already cited; confidence is recorded separately from performance in this pattern. - Role-playing exercises enhanced communication, collaboration and critical thinking [+W]: a review's one-line report of a role-play study with human partners, no effect size; consistent with the overall expectation of a gain, adds nothing to it. - Adult learners expect AI to support role play and assumption-challenging simulations [+W]: adult learners' expectations from interviews; useful when eliciting the learner-valued goal, not evidence about outcomes. - Participants with prior communication training scored higher on an AI-rated communication rubric [~W]: read 2026-10-19; an association in one uncontrolled study, useful for reading baseline differences between learners (the brief already records prior practice), not an effect of the practice. - Simulation and hybrid authentic-learning approaches do not necessarily differ in participants or experienced meaningfulness [~W]: read 2026-10-19; a journal editorial's summary of a CSCL taxonomy review, about collaboration settings rather than practice with a simulated person; no configuration follows. - Simulation is a site where professional knowledge is produced and reshaped, not only rehearsed [~W]: read 2026-10-19; an editorial's interpretive synthesis of qualitative studies (forestry, nursing, rescue, military); a framing, not a test. - A social-chatbot intervention was followed by lower loneliness and social anxiety [+W] (not about this page): read 2026-10-19; companionship for wellbeing, reported in a review, not rehearsal of a performance. - Student satisfaction with virtual patients is generally positive [+W]: read 2026-10-19; 17 satisfaction studies, most items indifferent between groups; satisfaction is recorded separately from performance here and does not change the model.

Illustrative design instance and observation record

Illustration only, invented, not a tested unit. An online, competency-based unit for working adults preparing to hold difficult conversations with clients: breaking unwelcome news and responding to the client's reaction. The designer's objective is that each learner, with a client they have not met before, acknowledges the reaction, checks understanding and agrees a next step, scored on a six-item rubric. One learner says she wants to stop over-explaining when people get upset. Session 1: learners watch a recorded strong and weak encounter scored against the rubric, then hold a 12-minute voice conversation with an LLM-played client whose disclosure rules were checked beforehand (it reveals a worry only if asked an open question). Each attempt is followed by item-level feedback from an automated assessor and a written debrief prompt; the learner then repeats the same conversation. Sessions 2–4 use new clients (an angry one, one who goes quiet, one who withholds a key fact), with fewer prompts and self-assessment before feedback; a mentor reviews one recording per learner and leads a 15-minute live debrief. Week 6: a blinded rating of a conversation with a human role-player the learner has not practised with; week 12: the same rubric applied to a real conversation, rated by a supervisor.

Record: performance and rubric → simulated person (who or what, disclosure rules, consistency check) → each learner's baseline encounter → feedback and debrief actually given after each attempt, by whom, and how soon → change between run-throughs of the same scenario → performance on new scenarios, by rubric item → anxiety and confidence, recorded separately → transfer rating with a different person at the stated horizon. Preserve the learner's stated purpose and any change in it. An agent may summarise this record but must not fabricate missing inputs or assign numerical state probabilities without a calibrated model.

Elements and limits

Simulation, role-playing, debrief, feedback, immediate feedback, worked examples, scaffolding, reflection and pedagogical agent signalling.

This pattern is scoped to rehearsal of an interaction with a person who responds. Procedural simulation with mannequins and part-task trainers, written cases with no one to talk to, simulations of physical systems, and assessment of competence for certification need distinct configurations and evidence. The wiki records no randomized comparison of an LLM role-play partner with a human standardized patient on a transferred communication outcome, no study of mixed-reality avatar classrooms for teacher candidates, and no study in business or technology settings of practice with simulated clients or customers, and no controlled study of an AI conversation partner for second-language speaking (only a review's thematic coding). The checks on the simulated person (content accuracy, push-back, fidelity of a simulated learner) rest on evaluations of the agents themselves, not on what learners gained from better-checked ones. The diagnostic branches and doses above remain to be tested with actual learners.

Examples

• Realism, control and feedback in teacher classroom-management training — the 3B virtual classroom with simulated pupils
• Text-based heterogeneity simulation with simulated peer feedback

Related Patterns

• Case-Based Learning
• Cognitive Apprenticeship
• Experiential Learning Cycle
• 4C/ID: Four-Component Instructional Design
• Competency-Based Learning
• Authentic Assessment
• Worked Examples
• AI Tutoring
• Peer Teaching
• Collaborative Learning

Key Sources

• Chernikova, O., Heitzmann, N., Stadler, M., Holzberger, D., Seidel, T., & Fischer, F. (2020). Simulation-Based Learning in Higher Education: A Meta-Analysis. Review of Educational Research. https://doi.org/10.3102/0034654320933544
• Cook, D. A., Hatala, R., Brydges, R., Zendejas, B., Szostek, J. H., Wang, A. T., Erwin, P. J., & Hamstra, S. J. (2011). Technology-Enhanced Simulation for Health Professions Education: A Systematic Review and Meta-analysis. JAMA, 306(9), 978–988. doi:10.1001/jama.2011.1234
• McGaghie, W. C., Issenberg, S. B., Cohen, E. R., Barsuk, J. H., & Wayne, D. B. (2011). Does Simulation-Based Medical Education With Deliberate Practice Yield Better Results Than Traditional Clinical Education? A Meta-Analytic Comparative Review of the Evidence. Academic Medicine, 86(6), 706–711. doi:10.1097/acm.0b013e318217e119
• Kononowicz, A. A., Woodham, L. A., Edelbring, S., Stathakarou, N., Davies, D., Saxena, N., Tudor Car, L., Carlstedt-Duke, J., Car, J., & Zary, N. (2019). Virtual Patient Simulations in Health Professions Education: Systematic Review and Meta-Analysis by the Digital Health Education Collaboration. Journal of Medical Internet Research. https://doi.org/10.2196/14676
• Kron, F. W., Fetters, M. D., Scerbo, M. W., White, C. B., Lypson, M. L., Padilla, M. A., et al. (2017). Using a computer simulation for teaching communication skills: A blinded multisite mixed methods randomized controlled trial. Patient Education and Counseling. https://doi.org/10.1016/j.pec.2016.10.024
• Yang, L., Liu, H., Li, S., Jia, R., Xiao, Y., Chen, G., & Lu, L. (2026). Evaluating Scaffolding-Oriented Multi-Agent Large Language Model System for Clinical Interview Training. doi:10.48550/arxiv.2609.10939
• Jiang, H., Wang, Z., Shen, W., Meng, M., Yang, D., Li, X., & Hao, Y. (2026). AI-Powered Simulation for Nursing Education: Mixed Methods Systematic Review. Journal of Medical Internet Research. https://doi.org/10.2196/95167
• Zhao, J., Yang, K., Gao, H., & Jiang, G. (2026). Generative artificial intelligence and the transformation of medical education: a scoping review. Frontiers in Education, 11, 1907589. https://doi.org/10.3389/feduc.2026.1907589
• Chen, A., Jin, S., Bao, C., Wang, C., Kraut, R. E., Wu, T., & Zhu, H. (2026). The Empirically Grounded Adaptive Virtual Patient for Psychotherapy Training: Disclosure That Responds to Therapist Micro-Skills. doi:10.48550/arxiv.2606.10051
• Hasan, M., Epstein, R., Carroll, T., & Hoque, E. (2026). Scalable AI-based clinical communication training and automated assessment. doi:10.48550/arxiv.2609.22517
• Johnson, W. L., & Lester, J. C. (2015). Face-to-Face Interaction with Pedagogical Agents, Twenty Years Later. International Journal of Artificial Intelligence in Education. https://doi.org/10.1007/s40593-015-0065-9
• Krämer, N. C., & Bente, G. (2010). Personalizing e-Learning. The Social Effects of Pedagogical Agents. Educational Psychology Review. https://doi.org/10.1007/s10648-010-9123-x
• Kobayashi, K. (2019). Learning by preparing-to-teach and teaching: A meta-analysis. Japanese Psychological Research, 61(3), 192–203. doi:10.1111/jpr.12221
• Chin, D. B., Dohmen, I. M., & Schwartz, D. L. (2013). Young Children Can Learn Scientific Reasoning with Teachable Agents. IEEE Transactions on Learning Technologies. https://doi.org/10.1109/TLT.2013.24
• Chen, A., Wei, Y., Le, H., & Zhang, Y. (2024). Learning-by-Teaching with ChatGPT: The Effect of Teachable ChatGPT Agent on Programming Education. arXiv. doi:10.48550/arxiv.2412.15226
• Yuan, Z., Xiao, Y., Li, M., Xuan, W., Tong, R., Diab, M., & Mitchell, T. (2026). Towards Valid Student Simulation with Large Language Models. doi:10.48550/arxiv.2601.05473
• Do, H., Sonkar, S., & Sachan, M. (2026). Simulating Students or Sycophantic Problem Solving? On Misconception Faithfulness of LLM Simulators. doi:10.48550/arxiv.2605.12748
• Vogel, F., Wecker, C., Kollar, I., & Fischer, F. (2016). Socio-Cognitive Scaffolding with Computer-Supported Collaboration Scripts: a Meta-Analysis. Educational Psychology Review (2017) 29, 477–511. https://doi.org/10.1007/s10648-016-9361-7
• Zhang, B., et al. (2026). A sequential integrating virtual simulation and authentic video for preschool teacher candidates' observational skills training: A design-based intervention study. PLoS One. https://doi.org/10.1371/journal.pone.0358951
• Wisniewski, B., Zierer, K., & Hattie, J. (2020). The power of feedback revisited: A meta-analysis of educational feedback research. Frontiers in Psychology, 10, 3087. doi:10.3389/fpsyg.2019.03087
• Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin, 119(2), 254–284. doi:10.1037/0033-2909.119.2.254
• Elliott, H., & Brumbaugh, K. (2021). Student Perceptions of a Simulated Clinical Experience: A Pilot Study. Teaching and Learning in Communication Sciences & Disorders, 5(1), Article 7. https://doi.org/10.30707/TLCSD5.1.1624982519.538001
• Kim, Y., Baylor, A. L., & PALS Group. (2006). Pedagogical Agents as Learning Companions: The Role of Agent Competency and Type of Interaction. Educational Technology Research and Development, 54(3), 223–243. https://doi.org/10.1007/s11423-006-8805-z
• Zhang, H., Qu, L., Zheng, J., et al. (2026). What Platform Scores Miss: Multidimensional Evaluation of AI Teaching Agents in Medical Education. JMIR Medical Education. https://doi.org/10.2196/96819
• Yang, K., Wang, C., Galley, M., Singh, C., Inala, J. P., Zhai, C., & Gao, J. (2026). StudentSim: Training LLM-based Student Simulators. doi:10.48550/arxiv.2609.01591
• Sitzmann, T. (2011). A Meta-Analytic Examination of the Instructional Effectiveness of Computer-Based Simulation Games. Personnel Psychology, 64, 489–528. doi:10.1111/j.1744-6570.2011.01190.x

Page metadata

Field	Value
type	pattern
id	practice-with-simulated-people
title	Practice with Simulated People
description	A reusable policy for rehearsing an interpersonal or professional performance with a simulated person (a standardized or virtual patient, a simulated client, customer or colleague, avatar or simulated students, a pedagogical agent, or an AI role-play partner): name the performance and how it will be judged, give novices an example first, run short repeated encounters with feedback and a debrief after each, vary and harden the scenarios across sessions, and judge the unit on a blinded performance with a different person rather than on in-simulation scores or satisfaction.
status	review
grain_size	unit
canonical	True
generated	by claude/unspecified on 2026-10-19
trust	unverified — no one has confirmed this page's content

Frontmatter is the page's machine-readable record — see the Schema & Guide.