AIEd Knowledge BaseAIEd KB
On this page

Concept

Simulation

created 2026-08-12 · updated 2026-11-10

Simulation — the use of modeled environments, agents, or scenarios to support learning through practice and feedback in contexts that are safe, repeatable, and often otherwise inaccessible. Simulations let learners act, make errors, and see consequences without real-world cost, and are increasingly powered by AI and agent-based modeling.

Questions to Consider

  • Recall a time you learned something by doing it in a safe, low-stakes environment — a lab, a mock exercise, a flight or game simulator. What made that practice effective, and what might be lost if the simulation were too realistic or not realistic enough?
  • The page argues simulations let learners make errors and see consequences 'without real-world cost.' What do you think is gained, and what might be lost, when the cost of a mistake drops to nearly zero?
  • If an AI can simulate patients, students, or conversation partners for practice, where would you draw the line between valuable rehearsal and practice that fails to transfer to real human interaction?
  • Why might a learner's awareness of a simulation's limits — its trustworthiness — matter as much as how faithfully it models reality?
  • How could the same simulation technology that helps someone learn also mislead them, and what would you need to know to tell those two outcomes apart?

Introduction

Simulation sits at the core of experiential and Active Learning pedagogies. It provides the deliberate practice, productive failure, and feedback loops that build skill and judgment. AI has transformed simulation in two ways: it powers more realistic and adaptive simulated environments, and it generates simulated learners, patients, or interlocutors that make practice scalable. Behavioral evidence shows that how learners engage with a simulation varies systematically rather than uniformly: tracing online learners building ecological models in VERA, An, Hammock & Goel (2025) classified engagement into Observation (frequent runs and parameter adjustment with little model building), Construction (hands-on building with little simulation), and Exploration (full construct–parameterize–simulate cycles), with Explorers producing the most complex and diverse models and observation-heavy learners largely copying existing ones — an argument for designing simulation environments that push learners toward full-cycle activity.

  • An explicit mechanism is inspectable in a way a learned generator is not. DagFlow-AI builds synthetic data from a directed acyclic graph plus per-variable distributions and formulas, so the data-generating process is a document an instructor can read and correct (Arthur et al., 2026). This is in deliberate contrast to GAN and VAE synthesis, which needs a seed dataset and yields an opaque process. The trade-off is that reaching a stable simulation took several iterations.

AI and simulation

  • AI-powered environments: adaptive simulations adjust difficulty and scenarios to a learner's state, linking to Adaptive Learning and Reinforcement Learning-based coaching.

  • The virtual environment and the AI layer are separable gains. In a three-week Grade 9 Ohm's law unit, adjusted achievement was 23.45 of 30 with an AI-supported virtual laboratory, 20.51 with the same laboratory without AI and 16.57 in a physical laboratory (Li and Ahmad (2026)). Intact classes were assigned with a different teacher each, so this describes three classroom arrangements, not the agent alone.

  • Prompt-generated simulations put a bespoke lab within reach. Instructors can generate browser-run models of the topic a course needs from a reusable prompt template (sliders, animation, time-dependent graphs), validating each twice - technically, then against the known analytical solution - instead of accepting the closest published simulation (Ben-Zion et al., 2025).

  • Automating the human counterpart. Chew et al. (2026) replace the specialist human role-players who staff air-traffic-control simulations with autonomous LLM sim-pilots, removing a training-capacity bottleneck; their fine-tuned speech pipeline cut word error rate on Singaporean-accented aviation speech from 107.80% to 23.45%, though all evaluations were component-level with no trainee cohort run.

  • Adversarial simulators as training partners. ResidencyRL (Liévin et al., 2026) pairs a policy agent with LLM patient simulators built to behave adversarially across a 57K-case scenario pipeline; training against them cut missed red-flag rates by about a third, and blinded clinicians preferred the trained agent in 87.6% of side-by-side comparisons.

  • Encode missing case states explicitly. A multi-agent standardized-patient system keeps intent recognition, case-grounded response generation and post-session evaluation separate, distinguishing four kinds of missing information — clinically negative, unknown to the patient, not yet assessed, and absent from the case — because collapsing them into "normal" injects unsupported clinical facts (Gao et al. (2026)).

  • Simulated agents: AI can simulate patients (for medical training), students (for teacher practice), or conversation partners, making high-stakes interpersonal practice accessible and repeatable. In teacher education, Zhuang and Zhang (2025) built Student GPT, a custom ChatGPT chatbot that role-played a middle school student holding common ratio-reasoning Misconceptions about AI, giving preservice mathematics teachers affordable, content-specific practice at diagnosing student thinking — and used an Affective, Communicative, Technical (ACT) coding framework to systematically assess the simulated student's role-play strengths (clarity, relevance, error consistency) and authenticity weaknesses (teacher-like tone, role confusion).

  • Make counterfactual history a first-class feature. SupplyNet's contextual multi-agent LLM agents generate supply-chain dynamics emergent from learner decisions rather than scripted, and its branching timeline lets learners revisit earlier choices without penalty — 13 of 14 participants rated it highly for connecting decisions to performance against 3 for the baseline (Li et al. (2026)).

  • Grounded dynamics, not a prompted persona. Chen et al. (2026) parameterized a virtual patient's disclosure dynamics from nearly 2,000 hours of real psychotherapy transcripts, then updated the level each turn; across 1,033 turns with 20 clinicians its disclosure rose with therapist empathy and exploration while a prompt-only baseline on the same LLM stayed flat.

  • Role-play puts the learner in the part. Where simulated agents supply the counterpart, role-play gives the learner that part instead. Sanoubari and colleagues (2026) had 18 children aged 9-10 watch a bullying scene enacted by social robots, reason about each character's position, then rehearse defending by puppeteering a robotic avatar, and reported gains in perceived Self-Efficacy for defending plus better-calibrated beliefs about whether confronting a bully actually stops it. Their framing, robot-mediated applied drama, keeps a human facilitator in the Forum Theatre role and confines automation to narrative control, which is a useful reminder that the demanding part of role-play is the reflection rather than the machinery. Lock, Arteaga and Johnson (2025) place role-play alongside simulation among the strategies that AI-supported online learning draws on.

  • Simulated learners: models of student behavior let researchers and designers test tutoring systems and curriculum before live deployment, grounding Learner Modeling and Adaptive Instruction and Knowledge Tracing.

  • Trust and fidelity: the value of a simulation depends on how faithfully it models the real context — and on the learner's awareness of its limits, connecting to Trust Calibration.

  • Human-likeness is a property of the interaction, not the model. Once both backends were rendered behind the same avatar, a text-based human-likeness advantage disappeared: participants called the commercial model natural (54% versus 22%) while communication performance was identical; voice, latency and turn-taking reshaped how the same dialogue quality was perceived (Hasan et al. (2026)).

  • GenAI in simulation-based learning. Neto and colleagues (2026) systematically review GenAI across scenario-, case-, problem-, and simulation-based learning in healthcare education, finding positive outcomes for higher-order cognitive skills but inconsistent results elsewhere, with hybrid human-AI collaboration outperforming fully automated approaches. Wenzel, Geiger, and Liening (2026) develop AI conversational agents for adaptive support in business simulation games, addressing the common gap of limited formative feedback and structured reflection in simulation-based learning.

  • The "authenticity gap" bounds what AI simulation can replace. In clinical simulation, Jiang et al. (2026)'s mixed-methods systematic review of AI-powered nursing simulation (19 studies, N=1,253) finds AI effective for cognitive knowledge and affective outcomes but inconsistent for complex psychomotor skills. They name an authenticity gap: a learner-perceived shortfall in emotional resonance, nonverbal cue recognition, and tactile/physical examination dimensions. That gap explains why AI simulation is best for highly structured objectives (foundational communication, history-taking) and should sit in a stepped simulation continuum that hands advanced psychomotor and emotionally complex scenarios to human-standardized patients and clinical placement. Technical instability (e.g., speech-recognition delays) can also add extraneous cognitive load and anxiety, so fidelity and stability are themselves design levers. This parallels Neto et al.'s finding that hybrid human–AI approaches outperform fully automated ones.

  • Teacher-AI co-designed simulations. Interactive simulations that support both conceptual learning and competency development are scarce in hands-on domains, and GenAI output often lacks pedagogical validity. In drone-based STEM education, teacher-AI co-designed simulations embedded in an otherwise identical hands-on curriculum were evaluated with a quasi-experimental pretest–posttest design across 30 secondary students, examining whether simulation-supported instruction yields superior learning outcomes (From simulation to flight: Simulation-assisted drone learning with teacher-AI co-designed scaffolds for secondary students' STEM knowledge and competencies). Separately, multi-agent tutoring benchmarks such as ASTRA use simulated socially intelligent agents to study participation-balanced collaboration in introductory programming (ASTRA: A synthetic benchmark for trace-based evaluation of socially intelligent multi-agent tutoring and participation-balanced collaboration in introductory programming).

  • Learner control in simulation is enacted, not granted. A 2 × 2 experiment in a flocking simulation (Su, Nair and Nagashima 2026) gave some students parameter sliders, some an optional conversational agent and some both; every condition improved, but neither affordance produced a reliable difference once prior knowledge was controlled (p = .849 and p = .108). What predicted gains was where and how long learners manipulated parameters: sustained slider use in the most conceptually complex lesson was positively associated with gains, and the same behavior in the easier lesson negatively. For simulation builders the implication is that offering controls is not the intervention — helping learners decide what to change, and register what changed, is.

  • Agency is a designed layer, not a property of the technology. Theelen et al. (2026) sort simulations by the practice represented, the simulated situation, how agency and adaptation are distributed, and how the learning is orchestrated, rather than by immersion level.

  • A digital twin is a simulation with a hardware bill, and the bill shapes who can take part. A systematic review of 11 studies of digital twins in engineering and STEM higher education found that every implementation delivered a working prototype with real-time synchronization, yet only three reported statistically significant learning gains. Physical rigs capped access at 1–3 concurrent students in five studies, and reduced social interaction was the most-cited pedagogical challenge, in six (Pelayo-González et al. (2026)).

  • Agent-based simulation can model a Metacognition theory as a testable system. An agent-based digital-twin-inspired framework combining the five-level AMERT model with a NetLogo simulation let researchers vary learner archetypes and task conditions and observe metacognitive trajectories, though the authors stress the output is hypothesis generation, not evidence about real learners (Guo, 2026). Simulation becomes a virtual laboratory for a theory that is otherwise hard to observe.

  • The curriculum is the unit that decides a simulation's value. The Curricular Bridge Model argues that return on simulation depends on how experiences are sequenced across a program — context reuse, variation in reasoning demands, evidence feeding progression — rather than on platform fidelity (Herrera Alcaíno, 2026).

  • Simulated clients outran the evidence. A scoping review of 14 sources on generative AI in counselor education found AI-simulated clients the dominant application, but no study measured trainee competency, used a comparison condition or tested crisis and risk-assessment scenarios (Generative AI in Counselor Education: A Scoping Review of Applications, Ethics, and CACREP Competency Mapping (2015-2026)).

Connections

Simulation connects to Active Learning, Adaptive Learning, and Pedagogical Agent. It is a mechanism for experiential and Constructivism learning and is amplified by AI's ability to generate adaptive, realistic practice environments.

Simulation-based learning now includes synthetic patients, students and role-play partners, and AI Avatars and Synthetic Instructors argues that their consent, disclosure and person-fidelity questions are not the same as equipment or scenario fidelity.

What cannot be delegated

Most work on simulated conversation asks how faithfully an agent plays its part. Álvarez-Rubio (2026) asks the prior question: what a teacher cannot hand over at all across the simulation cycle of briefing, scenario, and debrief. The paper separates four delegation operations: execution, conveyance, facilitation, and the undertaking of a commitment. Execution and conveyance let the teacher correct a failure afterward; facilitation hands over judgment about an activity's structure. The fourth fails differently, because a commitment's essential condition is answerability as a relation to those who may demand reasons. An agent can perform a debriefing function; it cannot be the party who answers for it. On this account a chatbot that debriefs as well as a human never becomes answerable, so stronger debriefing evidence cannot move the line.

The framework travels as a test rather than a catalog. Three questions decide any function: what would pass to the agent, whether the teacher could correct a failure, and who could be asked to stand behind the agent's words. A function with no answer to the third is not a candidate. Applied to preparation, the test yields six verdicts — diagnosis and logistics delegable, stated objectives partly delegable, and the fiction contract, psychological safety, and assessment terms refused. In the debrief the paper sets out three stances an agent might hold, expert, character, and Maieutic, and reports the Maieutic failure mode: an agent built to teach by never answering left students frustrated. It is a design framework, not an outcome study. The authors state that no data were collected for research, and the deployment reported is a pilot read through the teacher's own observations.

Common Misconceptions

  • "Simulation is cheaper and therefore second-best." Autonomous LLM sim-pilots replace the specialist human role-players who staff air-traffic-control simulations, removing a training-capacity bottleneck — the constraint was staffing, not fidelity. Where human role-players are scarce, the LLM simulation is not a substitute but the only available option.
  • "Simulators should behave like average cases." ResidencyRL (Liévin et al., 2026) pairs a policy agent with LLM patient simulators built to behave adversarially. A simulator designed to be cooperative makes the learner's practice easier than the clinic, which is the opposite of what training needs.
  • "Scripted scenarios teach decision-making under uncertainty." SupplyNet's contextual multi-agent agents generate dynamics emergent from learner decisions rather than scripted sequences. Scripted scenarios rehearse a known path; the emergent ones rehearse the decision.

Connected Concepts

Connected Articles

Connected Resources

  • OpenMAIC
    The open-source release of MAIC: a multi-agent classroom that turns a topic or document into slides, quizzes, interactive simulations and project-based activities, delivered by AI teachers and classmates.

Metadata

Learning and instructionActive Learning, Experiential Learning
Technologies and techniquesAdaptive Learning, Pedagogical Agent, Reinforcement Learning
SourceHistoryEditChat

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.