Integrating AI in Lesson Planning: Critical Reflection and Thoughtful Implementation

Synthesis: Wilcoxen, Steiner and Bell (2026) studied 26 beginning teachers in a university induction program who used generative AI to draft lesson plans, then documented every revision they made and taught a 20-minute version in a simulated classroom. Across 88 qualitative artifacts and 26 open-ended responses, they found that AI produced a usable procedural baseline — topics, objectives, activity lists — but systematically missed contextual and relational complexity. Teachers acted as pedagogical mediators: they swapped worksheets for movement-based tasks, rewrote language to match reading levels, added modeling and think-alouds, turned objectives into "I can" statements, re-sequenced lessons around classroom logistics, and built in relationship moves the plan never contained. The study is a small qualitative analysis, but it gives a concrete map of where AI lesson plans fail and what pedagogical judgment teachers supply.

Key Findings

1. AI produced a serviceable first draft, not a teachable plan. Content transformation was the least modified domain; topics and basic explanations were generally relevant and accurate.
2. Task design was consistently rewritten. Teachers replaced worksheets with hands-on sorting, gallery walks, movement-based grouping and simulations, noting that AI rarely explained how to group or manage an activity without substantial prompting.
3. Adaptation was the largest gap. Teachers added modeling, think-alouds and guiding questions for learner differences, and adjusted language to students' reading levels — the domain AI handled least well.
4. Clarity required expansion. Broad objectives became "I can" statements; directions were rewritten with examples, behavioral expectations and visual or auditory supports.
5. Sequencing and phasing were restructured. AI plans were linear and compartmentalized; teachers added material distribution, preparation steps, transitions and timing.
6. Teachers described their role as maintaining professional control. In survey responses they emphasized reviewing and revising outputs rather than accepting them, and named prompt clarity as a condition for useful output.

The study

Participants were 26 full-time beginning teachers in eight school districts across two states, most in their first year, teaching across PK–12 including special education, dual-language, music and art. They worked in an induction program where two of the authors were instructors; the authors acknowledge the power dynamic and report that study data were not analyzed until final grades were submitted.

The instructional cycle merged planning, revision, implementation and reflection. Teachers received a lesson-plan template, prompted an AI tool of their choice (ChatGPT, MagicSchool, Gemini and others were mentioned) to create a lesson for an assigned topic, then refined it into a 20-minute lesson with time stamps. They revised the plan collaboratively and recorded their reasoning for each change, taught the lesson to peers in a simulated classroom, and analyzed video against a rubric. The analysis used directed qualitative content analysis with a coding scheme from König et al.'s six cognitive domains of lesson planning: transforming content, creating and selecting tasks, adaptation, clarity, sequencing and phasing.

What the revisions show

The pattern across domains is consistent: AI was strongest where the task was to produce plausible content and weakest where the task required knowledge of particular students and classrooms.

• Transforming content was minimally altered. Teachers changed scenarios to reflect students' lived experiences and added community examples, treating relevance as their contribution rather than the model's.
• Creating and selecting tasks was heavily edited. Even when the age group was specified in the prompt, teachers redesigned activities for interaction and cognitive engagement, and supplied the grouping and management detail the plan lacked.
• Adaptation was the most prominent gap and the most significant teacher enhancement: support structures for learner differences, language simplification, and diagnostic awareness of who would need what.
• Clarity was expanded rather than corrected: objectives became student-facing statements, directions gained examples, and supports were added for visual and auditory learners.
• Sequencing and phasing were rebuilt around logistics — materials, preparation, transitions, timing — which the AI plan treated as implicit.

The teachers' own framing matters. They did not describe themselves as accepting or rejecting AI; they described reviewing, revising and customizing it to keep instruction sound and responsive. That is a more specific account of "teacher in the loop" than the usual phrase suggests: the loop is not approval but the addition of contextual knowledge.

What this means for practice

• Use AI for the procedural skeleton. Topics, objectives and activity ideas are a reasonable starting point; expect to rewrite tasks, supports and logistics.
• Prompt for context, but do not expect it to solve adaptation. Even explicit age and topic prompts did not produce the grouping and differentiation detail teachers needed.
• Make revision reasoning visible. Asking teachers to record why they changed each element turns plan revision into evidence of pedagogical reasoning, which is useful in induction and teacher preparation.
• Teach the mediation role explicitly. Beginning teachers benefited from naming what they supply that the model cannot: knowledge of students, classroom management, relationship-building and local examples.
• Pair AI planning with rehearsal. The simulated 20-minute lesson surfaced problems that plan review alone would not, and gave teachers evidence for reflection.

Limitations

• The study is small and qualitative: 26 beginning teachers in one induction program, with no comparison group and no measure of student learning.
• Lessons were taught to peers in a simulated classroom, not to PK–12 students, so claims about classroom effectiveness are indirect.
• Two authors were the participants' instructors, and the analysis relies on teachers' self-reported reasoning; the authors describe mitigation but cannot eliminate the power dynamic.
• The coding scheme was deductive and drawn from one framework, so domains outside König et al.'s six may be underreported.
• AI tools changed rapidly during and after the study, and the specific failure patterns may not hold for newer models.

Citation

Wilcoxen, C., Steiner, A., & Bell, J. (2026). Integrating AI in Lesson Planning: Critical Reflection and Thoughtful Implementation. The Interdisciplinary Journal for Research on Pedagogy and Learning Design, 1(1).