The Reflective Triangle Model: AI as a Cognitive Mediator in Teachers' Professional Learning and Learning-Community Development

Synthesis: This practice-based study, presented at the Vietnam Deeper Learning Conference 2026 and authored by a doctoral candidate at TU Dresden who is also a standing member of the Olympia Schools academic board, proposes the Reflective Triangle Model as a way to connect individual teacher reflection to shared professional knowledge inside learning communities. AI is framed as a cognitive mediator — a cognitive tool that surfaces evidence, prompts inquiry, and enables dialogue — rather than as an evaluator that judges teaching. The model integrates three interacting dimensions — self-reflection, collaborative reflection, and community reflection within a Professional Learning Community (PLC) — with AI mediating across all three levels. It is tested through a four-week illustrative case (N = 6 teachers at one school) aimed at feasibility rather than generalization, positioning AI as a supportive instrument for teacher AI competency and professional learning while preserving human professional judgment and Learner Agency.

The study begins from a practical problem: post-teaching reflection in schools is often memory-dependent, subjective, and easily truncated by workload, so it rarely becomes systematic professional learning. A second problem concerns the transfer of knowledge from the individual to the collective — personal insights do not automatically become shared professional knowledge in a PLC. In this context, the paper argues that generative AI need not only generate materials or automate tasks, but can genuinely participate in how teachers think about their own practice, positioning AI within a human-centered AI approach that supports rather than replaces teacher professional expertise.

The Reflective Triangle Model

The Reflective Triangle rests on three complementary theoretical foundations: professional reflection, collaborative learning in professional communities, and AI as a cognitive mediator. Professional reflection draws on the work of Schön (reflective practitioner, reflection-in-action versus reflection-on-action), Brookfield (critical reflection), and Argyris and Schön (single-loop versus double-loop learning). Learning communities build on Lave and Wenger's communities of practice and DuFour's PLC, with Timperley specifying the inquiry questions — not just "what did we do?" but "how are students learning?" and "what evidence shows it?" — that turn a PLC from a place of sharing experience into a space of professional inquiry.

The model itself integrates three interactive dimensions that form a continuous Feedback cycle rather than a linear sequence. Self-reflection begins from a classroom experience, with classroom evidence — transcripts, classroom discourse, questioning patterns, student engagement — as the object of reflection, and AI supporting analysis and questioning without supplying final conclusions. Collaborative reflection brings individual insights into the PLC, where AI becomes a shared mediating object for professional dialogue: members collectively examine data, verify claims, and propose alternative interpretations. Community reflection extends reflection from the group to the school's wider professional community, where insights built across many cycles can crystallize into shared principles, practices, and procedures — though whether an individual insight becomes organizational knowledge still depends on the community's social processes of verification and acceptance.

AI as a Cognitive Mediator

The paper distinguishes three levels of AI mediation in education. Task mediation changes what teachers produce — faster lesson plans, multiple-choice questions, rubrics — answering "can AI do this for me?" Evaluative mediation (AI-driven evaluation) changes judgments about practice — concluding whether a lesson was good — answering "how does AI judge me?" Epistemic mediation neither produces artifacts nor judges practice; instead it changes how teachers approach, organize, and verify evidence so they can reach their own understanding, shifting the question from "what does AI say?" to "what does this evidence make me ask?"

The Reflective Triangle locates AI at the epistemic level. Building on Vygotsky's account of mediation, the paper stresses that epistemic mediation is not neutral: by choosing which indicators to surface (e.g., teacher talking time) and which framework to use (e.g., Bloom's taxonomy), AI frames what teachers can see and therefore question. AI's three defined functions are evidence surfacing (making hard-to-see patterns visible), inquiry prompting (generating questions and alternative interpretations), and dialogic mediation (supplying input for dialogue with oneself and with colleagues in the PLC). Crucially, AI-generated analysis is not treated as objective evidence, and the model explicitly contrasts AI-supported reflection with AI-driven evaluation, preserving professional judgment as human-led.

The Practice-Based Case at Olympia Schools

The model was examined through an illustrative practice case at one school — N = 6 teachers over four weeks — designed to investigate feasibility in a real context rather than to generalize effectiveness. Three research questions guided the study: how AI supports teacher reflection when used as a cognitive mediator (RQ1); whether bringing AI-assisted evidence into the PLC changes the quality of professional dialogue (RQ2); and what conditions, tensions, and limits arise in implementation, including teacher agency, dependence on AI, data quality, psychological safety, and Privacy (RQ3). Consistent with a practice-based thematic analysis, the aim was systematic learning from a designed practice in a specific context rather than a claim of universal efficacy.

The study's contribution lies not in a new technical AI capability but in an architecture of connection: it explicitly describes the mechanism by which an AI-assisted insight surfaced at the individual level moves into collaborative dialogue and is then normalized into community-level knowledge in a continuous feedback loop, rather than three separate applications. The paper situates this against prior work on AI-mediated reflection (mostly operating at a single level) and AI-enhanced PLCs (which assume data already exists without explaining how it is generated and interpreted individually before reaching the PLC).

What this means for practice

• Instructors. Reflect on captured classroom evidence — transcripts, classroom discourse, questioning patterns, student engagement — rather than memory, since the study's starting problem is that post-teaching reflection is memory-dependent and easily truncated by workload.
• Instructors. Keep AI at the epistemic level and out of judgment about practice: shift the question from "what does AI say?" to "what does this evidence make me ask?", and do not let the system conclude whether a lesson was good.
• Faculty developers. Route individual AI-assisted insights into the PLC for collective verification and alternative interpretation before treating them as professional knowledge; the model's contribution is the mechanism by which an individual insight is normalized into community-level knowledge.
• Faculty developers. Choose indicators and frameworks deliberately, because AI frames what teachers can see by selecting which indicators to surface (for example, teacher talking time) and which framework to apply (for example, Bloom's taxonomy).
• Faculty developers. Protect psychological safety and teacher Learner Agency: the paper reports that AI supports psychological safety only when its use is embedded within a non-punitive PLC, and that AI-generated analysis must not be treated as objective evidence. Judge AI's value by how far it supports teachers' own capacity and autonomy rather than by task efficiency alone, since AI-generated analysis depends on input data, prompting, model and user interpretation, and the model therefore requires organizational conditions that protect professional responsibility and govern how reflection cycles, data selection and peer verification of AI-assisted claims are organized.

Limitations

• The evidence comes from one practice case of N = 6 teachers at a single school over four weeks, with one professional group and no control group or statistical testing; the authors state the figures illustrate the direction of change and cannot be statistically generalized to other schools or education systems.
• Psychological safety was measured only through semi-structured interviews conducted at the end of Week 4 and reported retrospectively, which the authors flag as carrying recall bias and as a qualitative indicator rather than an independent two-time-point measurement.
• The tool under study also generated part of the research evidence: AI-supported analysis is not independent evidence and must be corroborated across multiple data sources, and the authors note the absence of reported inter-rater reliability among independent human coders.
• The design cannot isolate AI's specific contribution from PLC structure, peer support, teachers' own reflective capacity, or school culture, so the findings are evidence of feasibility, observed change, and emerging mechanisms rather than of causal relationships.

Citation

Nguyen, T. T. T. (2026). The Reflective Triangle Model: AI as a Cognitive Mediator in Teachers' Professional Learning and Learning-Community Development. EdArXiv preprint.