Analyzing teacher-AI interaction patterns across teacher experience and AI proficiency in student-centered lesson design

Synthesis: Choi, Jeong, Park, Kim, and Han (2026) analyze how South Korean elementary teachers interact with ChatGPT while designing student-centered lessons, using CORDTRA diagrams of ten teachers' think-aloud planning sessions to compare interaction patterns across teaching experience and AI proficiency. Seven distinct interaction patterns emerge — direct adoption, elaborated adoption, initial rejection, revised adoption, follow-up guided use, complex interactions, and bypassing AI — and they distribute differently by case. Experienced teachers with high AI proficiency critically adapt AI output to classroom context through re-prompting and elaboration, whereas less experienced or less AI-proficient teachers rely more heavily on AI suggestions and engage in less contextual adaptation. The findings position generative AI as a collaborative co-designer whose value depends on the interplay of teaching experience and AI proficiency, and argue for differentiated support to foster teachers' critical and reflective use of AI.

Key Findings

Seven teacher-AI interaction patterns. Analysis of 82 cord boxes from ten elementary teachers' ChatGPT-assisted lesson planning revealed seven predominant patterns: direct adoption (Initial Prompting → Accepting AI Response → Lesson Planning), elaborated adoption (adding refinement before use), initial rejection, revised adoption (rejecting then re-prompting), follow-up guided use, complex interactions (mixed acceptance and rejection), and bypassing AI (planning from textbooks or other resources instead). The two most common — elaborated adoption (29%) and direct adoption (21%) — show teachers largely either revising or directly using AI responses; complex interactions (15%) reflect selective incorporation.

Experience × AI proficiency shapes interaction. Teachers were grouped into four cases crossing experience (high ≥7 yrs / low ≤3 yrs, per Huberman's stages) with AI proficiency (high/low on Celik's Intelligent-Technological Pedagogical Content Knowledge (TPACK) survey). Experienced, AI-proficient teachers most often used elaborated adoption (33%), critically examining and refining AI output against educational context and frequently "considering students and context." Experienced teachers with low AI proficiency showed fewer interactions and, dissatisfied with AI output, tended to rely on their own expertise or textbooks rather than re-prompting. Less experienced teachers — regardless of AI proficiency — showed high acceptance (direct adoption predominant) and rarely "considered students and context," treating AI as a content provider or near-peer rather than critically adapting outputs.

AI proficiency versus pedagogical experience diverge. High AI proficiency alone did not produce critical use: low-experience/high-AI-proficiency teachers accepted responses with minimal modification (Patterns 1/2/5 ≈ acceptance), while experienced teachers grounded adaptation in pedagogical expertise. AI proficiency enabled technical communication; teaching experience supplied the contextual judgment to evaluate outputs. Teacher quotes illustrate the gap — a novice called AI "a coworker or a collaborative teacher," while an experienced low-proficiency teacher struggled to phrase prompts precisely.

Distributed-cognition framing. The authors interpret results through Distributed Cognition Theory (Hutchins), framing generative AI as a cognitive artifact that extends teachers' capacity to store, retrieve, and manipulate instructional knowledge. For novices, cognitive load is largely delegated to the system — an AI-dominant distribution; for experienced, AI-proficient teachers, an "optimal" distribution emerges where expertise and AI complement each other. Because GenAI actively generates and co-constructs rather than merely stores, they recast this as participatory shared cognition, echoing sociocultural co-construction (Vygotskian dialogue and joint activity) and positioning AI as a co-designer while the teacher remains the pedagogical decision-maker.

Need for differentiated teacher support. Implications center on matching support to the experience × proficiency profile: response-evaluation checklists and prompt templates for novices (to build critical adaptation and contextual judgment), and hands-on skill-building professional development for experienced teachers with lower AI proficiency. Collaborative professional development and peer mentoring are suggested, alongside ethical reflection and attention to teacher Learner Agency, Trust, and data-privacy concerns.

What this means for practice

• Instructors. Evaluate every AI suggestion against your students and classroom context before adopting it: the most common pattern among experienced, AI-proficient teachers was elaborated adoption (33%) — examining and refining output — not acceptance.
• Instructors. Diagnose the interaction pattern you default to and treat rejection as the start of a re-prompting cycle rather than the end of the exchange; less experienced teachers showed predominantly direct adoption and rarely considered students and context.
• Faculty developers. Build development around pedagogical judgment rather than tool fluency alone: AI proficiency enabled technical communication, but teaching experience supplied the contextual judgment to evaluate outputs, so pair response-evaluation checklists and prompt templates for novices with hands-on prompting practice for experienced teachers who lack AI proficiency.
• Faculty developers. Pair teachers across experience and AI-proficiency profiles for peer mentoring and collaborative professional development, so experienced teachers' contextual judgment and less experienced teachers' technical fluency transfer in both directions.

Limitations

• Ten elementary teachers in South Korea: the authors state the sample size was relatively small and the elementary-level focus limits generalizability to other educational levels and subjects.
• Sorting participants into four cases by experience and AI proficiency yields small cells, and the authors report that this design does not support robust comparative conclusions — findings should be interpreted with caution.
• Data come from think-aloud planning sessions with ChatGPT (82 cord boxes) rather than enacted classroom lessons, and the study did not evaluate whether the AI-assisted lesson plans actually reflected student-centered instructional principles.

Citation

Choi, S., Jeong, S., Park, S., Kim, Y., & Han, I. (2026). Analyzing teacher-AI interaction patterns across teacher experience and AI proficiency in student-centered lesson design. Teaching and Teacher Education, 169, 105266.