Classroom simulations for social-emotional learning and inclusive teacher education: a four-layer conceptual framework in the age of AI

Synthesis: Theelen and colleagues define a classroom simulation as a deliberately designed approximation of classroom practice in which a teacher takes a professional role, interprets classroom information, and acts or decides in interaction with represented students, parents, or colleagues, all inside a structured cycle of preparation, enactment, and debriefing. Their argument is that the field's habit of sorting simulations by technology or immersion hides the differences that matter, so they propose four interdependent layers: the professional practice represented, the design of the simulated situation, the distribution of Learner Agency and adaptation, and the orchestration of learning. Generative AI, on this account, is not a new category of simulation but a force that moves authorship and control across the layers. Educational value depends on alignment among the four, with inclusion, emotional safety, and human oversight treated as design responsibilities.

Conceptual paper: the authors synthesize prior literature into a framework and five theoretical propositions. No data were collected or analyzed, so the propositions are expectations awaiting empirical test, and every figure below is a count of the framework's own components or an attributed claim about another study, not a finding of this one.

Key Findings

1. Technology does not define the genre. Peer role play, actor-based clinical simulation, branching scenarios, intercultural virtual exchange, video, mixed-reality classrooms, virtual reality, serious games, and AI-supported encounters are all called simulations. Sorting them by modality, and treating immersion as progress, misreads the field: a highly immersive environment may offer little meaningful interaction, while a seminar-room role play can carry high psychological and functional fidelity.
2. The definition rests on role and process. The situation is deliberately designed for a learning purpose; participants take a professional role and must interpret, decide, act, or interact rather than only observe; represented actors include parents and colleagues as well as students; modality is a variable; and enactment is one phase among preparation, briefing, and debriefing.
3. Functional and psychological fidelity outrank realism. What matters is whether perception, action, response, and consequence correspond to the professional task. Purposeful simplification, such as fewer students or one high-leverage practice, is a design decision rather than a deficiency, and the access, prominence, simplicity, and clarity of diagnostic information give designers a principled way to raise complexity beyond adding avatars.
4. Four layers answer four questions. Professional practice asks what is practiced; the simulated situation asks how it is represented; agency and adaptation ask who or what shapes the interaction; orchestration asks how learning is facilitated. The layers are analytical rather than chronological, nested as increasing pedagogical specification, and a decision at any one constrains the others.
5. Agency is not representation. An avatar is what a participant sees; the actor behind it may be a peer, a simulation specialist, a script, a rule-based model, or an AI agent, so two identical virtual students can offer different encounters. Distribution of agency also defines an inclusive simulation: students who can only wait to be managed reproduce the hierarchy the design claims to question.
6. AI relocates authorship. Scripted designs fix the paths before use; agentic designs specify goals, constraints, profiles, memory, and decision policies, and the interaction emerges during use. The authors separate simulated social agents, coaching agents, and orchestration agents, and insist that the coaching role must not be blurred with the simulated student.

Why the technology labels fail

Hybrid design already separates modality from agency: a virtual environment may preserve direct human interaction, a digitally represented character may be run by a human operator, and video supports learning only when it is folded into facilitated reflection. Generative AI sharpens the problem, because large language models and multi-agent systems can generate open dialogue, hold character profiles, and coordinate several simulated actors at once, none of which a non-immersive versus immersive distinction describes.

The framework is positioned against four traditions: fidelity frameworks, which explain validity but say little about pedagogy; practice-based teacher education, which identifies what to rehearse but under-specifies architecture; AI simulation frameworks, which explain responsiveness but treat debriefing as background; and debriefing frameworks, which explain how experience becomes learning but treat design as background. The claimed contribution is integration rather than new constructs. Each layer has been discussed before, and what is new is treating them as one mutually constraining system.

The four layers and the questions they answer

Layer one is professional practice, and it comes first so that technology cannot set the learning goal. A difficult parent conference needs open dialogue and emotional responsiveness, not a visually detailed classroom; noticing participation patterns needs salience and attention distribution; inclusive adaptation needs students' thinking made visible without reducing them to diagnostic labels.

Layer two is the simulated situation: environment, actors, available information, permitted activities, and required decisions. Its key property is an internal logic: a scenario can invite one correct response or sustain several defensible courses of action. For Social-Emotional Learning and inclusive practice, the authors flag over-simplification into right and wrong choices as a specific risk, because interpersonal situations rarely have one universally appropriate response, and they demand that difference be multidimensional rather than a deterministic variable or a flat caricature.

Layer three is agency and adaptation, where the paper is sharpest. Beyond who controls the interaction, adaptation has to be pedagogically meaningful rather than merely variable, and Chernikova and colleagues' meta-analysis is used to distinguish adaptivity, where the system decides, from adaptability, where the learner decides. Its finding that computer-controlled adaptivity looks promising for Scaffolding while learner control may suit task progression is turned into an argument against treating full automation as the ideal.

Layer four is learning orchestration, the coordination of mediation across roles, phases, and tools: actors create consequential moments, facilitators organize their interpretation, peers widen the range of perspectives, and video anchors reflection in observable evidence. Debriefing carries particular weight for social and emotional practice, and orchestration must include emotional as well as cognitive scaffolding. Discomfort can unsettle habitual thinking, but the authors insist it is not automatically transformative and needs a psychologically safe container, paced briefing, and structured reflection.

What generative AI moves, and what stays human

Generative systems do not add a simulation category; they redistribute authorship among the teacher educator, the prospective teacher, and the computational system. The design problem shifts from writing every line to governing a space of possible behaviors, which buys variability and responsiveness at the cost of predictability. Emerging evaluations already report logical inconsistency, implausible attentiveness, and weak emotional and linguistic authenticity in LLM-simulated learners. The authors conclude that fluency and a sense of presence are not validity evidence.

Three consequences follow. Adaptation must be transparent, with learners and teacher educators knowing what information drives a decision and which parts of the simulation are controlled by people, scripts, or AI; human-in-the-loop design should be treated as a principled distribution of expertise rather than a stage on the way to automation. Multimodal signals such as voice, gaze, and movement may extend reflection, but nonverbal meaning is context dependent, automated affect recognition remains contested, and measurement can undermine psychological safety by implying that competence is a behavioral score. Personalization raises questions the framework cannot settle alone: which learner characteristics are pedagogically relevant, who sets the desired challenge, and whether participants can see and contest the model behind an adaptation.

Inclusion as a design property

Inclusion is given operational criteria rather than appended as a value. An inclusive simulation represents diversity as intersectional and embedded in complex social contexts; distributes agency so that marginalized or neurodivergent characters can initiate, resist, and negotiate meaning; and requires prospective teachers to confront inequity, bias, and power in real time rather than observe them from outside. The ethical corollary follows: AI systems have historically underrepresented disability experience, simulated students are designed representations rather than neutral reproductions of children, and a fluent generated response can look authentic while resting on inaccurate, culturally narrow, or stigmatizing associations. The remedy is participatory, with input from the communities portrayed and debriefing that asks whose perspective defines successful performance.

What this means for practice

• Instructors. Choose the professional practice before the platform. Name the teaching move or dilemma you want rehearsed, and judge a simulation by whether its responses and consequences correspond to that practice rather than by how realistic it looks.
• Instructors. Protect the debrief. The paper's clearest claim is that experience alone teaches little, so time to reconstruct perceptions, surface assumptions, and rehearse alternatives is part of the intervention.
• Instructional designers. Make control visible. Show who or what drives each response and what evidence an adaptive decision rests on, and leave learners and teacher educators able to decline a recommendation.
• Instructional designers. Write scenarios in which difference is multidimensional, avoid profiles that link a background characteristic to a behavior, and build in moments where a simulated student pushes back on the teacher's assumptions.
• Administrators. Fund the human side of the apparatus. Facilitation, actor training, and community review of scenarios are what make an AI-enabled simulation defensible, and they scale less cheaply than the software.

Limitations

• The framework is a conceptual synthesis with no empirical test: the authors state that its analytical utility has not been examined, and the five propositions are expectations rather than findings.
• The layers are conceptually distinct but functionally overlapping; the authors concede that the boundary between dynamic agency and adaptive orchestration blurs in highly automated environments.
• The literature synthesized is predominantly Western, so its assumptions about inclusive pedagogy, emotional expression, and psychological safety may not transfer without contextual adaptation.
• No search strategy, inclusion criteria, or synthesis protocol is reported, so the derivation of the four layers cannot be audited as a review, and the claim that they are non-redundant rests on the author team's collaborative judgment.
• The treatment of AI rests largely on adjacent pilot and evaluation work rather than validated models of simulated learner behavior, so its statements about agentic simulations are projections about a design space still being built.

Citation

Theelen, H., Angelini, M. L., Levin, O., Hemi, M. E., Kasperski, R., Dieker, L., Ledger, S., Siker, J., Dotger, B., Chernikova, O., Ye, J., Bradbury-van Heel, M., Chimicz, D., Dexter, S., Flavian, H., Lewicka-Zelent, A., Marc, A., Pop, R., & Torrijos Marti, I. (2026). Classroom simulations for social-emotional learning and inclusive teacher education: a four-layer conceptual framework in the age of AI. Education Sciences, 16, 1813.