Pedagogical resilience reconstruction in the AI era: a longitudinal reflexive thematic analysis of teacher reflective journals

Objective Generative artificial intelligence (GenAI) lowers the threshold for producing work, not necessarily for learning. This study examines how an instructor kept an interdisciplinary AI course teachable for first-year art students, where that effort reached its limits, and how the instructor interpreted students' engagement with AI tools. Methods The study is a longitudinal practitioner self-study of a 12-week course taught in autumn 2025 at a higher vocational college in northwestern China. The data are the instructor's own reflective journals: structured entries written during the semester for Weeks 1–7 and retrospective accounts of Weeks 8–12 written after the semester ended. The two strata are reported separately. Reflexive thematic analysis produced 40 evidence anchors, five themes, one cross-cutting thread, and one central organizing theme, with a full audit trail. Results The instructor's repairs (translating content into everyday and discipline-specific tasks, working around broken or restricted lab equipment, and demonstrating on the teacher's computer) kept the course running. However, as content moved toward frontier tools that required paid accounts, computing resources, and prior AI experience, teaching drifted toward demonstration (“the more cutting-edge something is, the more it can only be demonstrated”). Across the semester the instructor recorded a recurring pattern, here termed a teacher-perceived low-threshold illusion , in which students appeared, in the instructor's account, to treat seeing, bookmarking, or one-click generation as equivalent to being able to do. The instructor associated this pattern with three co-occurring conditions: natural-language prompting and one-click generation, demonstration-only teaching, and a final assessment that students could complete without having attended. Participation could be high while, in the instructor's judgement, students gained little. Conclusion Pedagogical resilience reconstruction in AI teaching is bounded. Keeping an AI course running did not by itself make learning visible: the journals record a tension between engagement, product completion, and evidence of learning, not an established failure of learning. In-course repair reached limits shaped by access conditions and an inherited assessment format, and reconstruction of the course took place at the scale of the semester. We present pedagogical resilience reconstruction as an emerging construct for multi-case testing.