AI Lesson Plans Pass the Time Test but Fail the Classroom Test, Review Finds

Artificial intelligence has swept into classrooms with astonishing speed, and nowhere is that more visible than in the quiet, labor-intensive craft of lesson planning. Teachers around the world now routinely ask chatbots to draft objectives, design activities, and generate assessment questions in seconds. But can a large language model actually produce a lesson plan that is pedagogically sound and practically usable in a real classroom? A new narrative review published in Frontiers of Digital Education by Vassilis A. Failadis, Sotiris K. Tasoulis, Spiros V. Georgakopoulos, and Vassilis P. Plagianakos of the University of Thessaly takes the most rigorous look yet at that question, synthesizing 53 recent publications focused specifically on lesson planning rather than on artificial intelligence in education at large.

The review is notable for its methodological discipline. Where much of the existing literature takes a broad, sweeping approach to AI in schools, the Greek team narrowed its scope deliberately to the underexplored niche of lesson planning. The analysis follows the search–appraisal–synthesis–analysis framework, a structured procedure for identifying relevant studies, judging their quality, and extracting coherent findings from a heterogeneous body of work. The 53 studies were then organized around four central themes: the quality and pedagogical value of AI-generated plans, the challenges and limitations of the technology, teachers’ attitudes and perceptions, and proposed improvements and future directions. This thematic structure allows the authors to draw a clearer picture than previous, more diffuse surveys of the field.

The headline finding is a paradox that will resonate with any teacher who has experimented with ChatGPT, Google Gemini, or Microsoft Copilot. Large language models are genuinely good at helping educators organize content and save time. They can produce a complete, well-formatted lesson plan in seconds, complete with learning objectives, staged activities, and differentiation suggestions, a task that might otherwise consume an evening. Yet the literature consistently shows that this output often lacks the flexibility, contextual awareness, and pedagogical depth needed for actual classroom use. The plans look right on the page; whether they work with thirty restless adolescents on a Friday afternoon is another matter entirely.

The technical reason for this shortfall lies in how large language models work. These systems generate text by predicting likely continuations based on patterns in their training data, not by drawing on lived teaching experience or knowledge of a specific group of learners. The result, documented across the reviewed studies, is output that tends toward the generic. A model asked to plan a physics lesson for junior high school students cannot know that this particular class has already struggled with a prerequisite concept, or that the school lacks the equipment a hands-on activity requires. Studies cited in the review, including comparative analyses of ChatGPT and Gemini in physics education and content analyses of AI-generated mathematics plans, repeatedly identify this absence of situational judgment as the technology’s central weakness.

Contextual blindness extends to culture and history as well. Research on foreign language lesson plan creation has documented trends, variability, and historical biases in chatbot output, suggesting that the plans a model produces may quietly encode assumptions about curriculum, pedagogy, and even politics drawn from skewed training data. Work in Japanese teacher education has examined the social, cultural, and political dimensions of relying on generative AI in lesson planning, while studies of AI-generated English lesson plans for students with intellectual disabilities raise pointed questions about whether these tools can serve learners with specialized needs. A lesson plan is never a neutral document, and the review makes clear that outsourcing its drafting does not outsource its values.

Perhaps the most striking strand of evidence concerns what happens when humans and machines are compared directly. In one study of music education, teachers were asked to label lesson plans as AI-generated or human-made and to rate their quality, with results revealing how difficult that discrimination has become and how quality judgments vary. Other experimental work has asked whether lesson plans created by ChatGPT are more effective than teacher-designed ones, with mixed outcomes. Meanwhile, studies applying frameworks such as Universal Design for Learning to AI-generated plans have found the output serviceable but shallow, prompting one research team to title its analysis with the pointed question of whether AI-generated plans are better than nothing. The overall picture is of tools that clear a low bar convincingly but have not yet demonstrated superiority over experienced professionals.

Teachers’ own attitudes, the review’s third theme, are more nuanced than either hype or panic would suggest. Across studies of pre-service and in-service teachers in mathematics, science, English language teaching, social studies, and vocational education, educators consistently acknowledge the time savings and organizational help that chatbots provide. Many teacher education programs have begun integrating these tools directly, using ChatGPT as a reflection tool to promote the lesson planning competencies of trainee teachers, or as an assistant with which preservice secondary mathematics teachers rehearse their planning. Yet the same studies emphasize, almost without exception, the importance of teacher involvement in reviewing, adapting, and critically evaluating whatever the model produces. The consensus position is not replacement but supervision.

The review also documents a rapidly evolving effort to make the tools themselves better. Computer scientists have proposed systems such as LessonPlanner, designed to help novice teachers produce pedagogy-driven plans with large language models, and generation pipelines that use self-critique prompting, in which the model evaluates and revises its own draft plans. Retrieval augmented generation has been prototyped for Ugandan secondary schools facing a new national curriculum, grounding model output in authoritative local documents rather than relying on parametric memory alone. Fine-tuned assistants tailored specifically to lesson planning have been tested with teachers to compare perceptions and use. These engineering approaches attack the technology’s weaknesses directly, though the review suggests they remain early-stage.

Against these advances, the authors identify a significant shortfall in the existing literature: there is still limited empirical evidence directly comparing lesson plans designed by teachers with those produced by large language models in controlled settings. Much of the 53-study corpus relies on perceptions, small case studies, or qualitative analysis rather than head-to-head experimental comparisons measuring student outcomes. Until such evidence accumulates, claims that AI can plan lessons as well as or better than teachers rest on shaky ground. The review’s authors argue that as these tools become more common, the field needs practical research, thoughtful integration into teaching practice, and ongoing professional development to ensure responsible and effective use in education.

The broader lesson of the review is about the division of labor between human expertise and machine fluency. Large language models excel at producing well-structured text quickly, and lesson plans are, on their surface, exactly that. But effective teaching depends on knowing learners, adapting in real time, and exercising professional judgment, capacities that no current model possesses. The literature synthesized by the University of Thessaly team points toward a future in which AI handles the scaffolding of planning while teachers supply the pedagogical depth, contextual sensitivity, and critical oversight that turn a plausible document into a good lesson. That future is not automatic; it depends on the empirical research and teacher training the review calls for, and on educators who treat every machine-drafted plan as a first draft to be interrogated rather than a finished product to be delivered.

Subject of Research: The role of large language models in generating and evaluating lesson plans in education

Article Title: Evaluating the Role of Large Language Models in Lesson Planning: Insights from a Narrative Review

Article References: Failadis, V. A., Tasoulis, S. K., Georgakopoulos, S. V., & Plagianakos, V. P. (2026). Evaluating the Role of Large Language Models in Lesson Planning: Insights from a Narrative Review. Frontiers of Digital Education, 3(3), Article 25. https://doi.org/10.1007/s44366-026-0099-6

Image Credits: AI Generated

DOI: 10.1007/s44366-026-0099-6

Keywords: large language models, lesson planning, ChatGPT, generative AI, teacher perceptions, pedagogical quality, narrative review, teacher education, AI in education, instructional design, retrieval augmented generation, pre-service teachers