Connecting Feedback to Choice: Understanding Educator Preferences in GenAI vs. Human-Created Lesson Plans in K-12 Education -- A Comparative Analysis

AI-Powered Lesson Plan Generation

Context: K-12 math education

Implementation: Utilizes AI models (GPT-4, LLaMA-2-13b) for lesson plan creation. Implementation involves prompt engineering (for GPT-4) and fine-tuning on expert-designed lesson plans (for LLaMA-2-13b).

Outcomes: Demonstrated competitiveness in cool-down activities and structured learning activities, particularly in high school settings. Excelled in content depth and structured content delivery at the high school level.

Challenges: Less favored than human-created plans overall, especially in elementary education. Lacked the nuanced differentiation, real-world contextualization, and student discourse facilitation of human-authored plans. Fine-tuned models were also slightly less favored than human-created content overall.

Human-Created Lesson Plans

Context: K-12 math education

Implementation: Lesson plans created by experienced curriculum designers.

Outcomes: Generally favored, especially in elementary education. Valued for nuanced differentiation, real-world contextualization, student discourse facilitation, and adaptability.

Challenges: Time-consuming and cognitively demanding to create. Early-career educators face considerable challenges in designing lesson plans that balance content rigor, pedagogical effectiveness, and classroom adaptability.

Lack of domain-specific knowledge and factual accuracy

GenAI models lack domain-specific pedagogical knowledge, leading to content that may be factually incorrect, pedagogically misaligned, or culturally insensitive. Educators report issues such as insufficient depth in subject matter and frequent errors in STEM-related content.

Proposed Solutions: Fine-tuning models with domain-specific datasets and incorporating education feedback loops to enhance AI-generated lesson plans. Integrating AI with professional curriculum design expertise. Fine-tuning models for subject-specific content.

Unpredictability and lack of adherence to educational standards

The unpredictability of AI-generated lesson plans raises concerns about their reliability, relevance, and adherence to educational standards.

Proposed Solutions: Rigorous evaluation of AI-generated outputs to ensure pedagogical soundness and contextual relevance. Refining GenAI models to better align with teacher preferences by incorporating user engagement data into model training.

Lack of nuanced differentiation, real-world contextualization, and student discourse facilitation

AI-generated plans often lack the nuanced differentiation, real-world contextualization, and student discourse facilitation of human-authored plans, particularly in elementary education.

Proposed Solutions: Refining GenAI models to incorporate user engagement data into model training, emphasizing elements such as scaffolded support, differentiated instruction, and interactive learning strategies.

Generalizability limitations

The study's findings may not be universally applicable due to the geographic, demographic, and contextual diversity of the educator sample. Pedagogical priorities may differ across regions, school types, and educational systems, which could influence educator preferences for AI-generated content. Educator biases, familiarity with AI tools, and subjective experiences may have shaped their evaluations.

Proposed Solutions: Expand participant diversity across regions, subjects, and teaching styles. Conduct longitudinal studies to track AI’s long-term impact on student learning outcomes.

Project Contact: Dr. Jianhua Yang

LLM Model Version: gemini-2.0-flash-lite