Performance Evaluation of Generative AI Models in Chemistry Lesson Design Using Role-Based Prompting: Insights from Micro- and Nanoplastics

Recent advances in generative artificial intelligence (GenAI) offer new possibilities for supporting teachers in instructional design; however, comparative empirical evidence on the performance of different models in discipline-specific lesson planning remains limited. This study evaluates five widely accessible GenAI models—ChatGPT (GPT-5.2), Claude (Sonnet 4.5), DeepSeek, Google Gemini 1.5, and Microsoft Copilot—in generating secondary-level chemistry lesson plans using role-assigned prompting (RAP). The instructional context was framed around the socio-scientific issue (SSI) of microplastic and nanoplastic impacts.

All models received the same role-assigned prompt (RAP) to act as experienced chemistry teachers and design curriculum-aligned lesson plans for 11th-grade students, which were then evaluated by five veteran chemistry teachers using an analytic rubric. The author-developed rubric, based on relevant theory, evaluated eight criteria, including learning outcome alignment, 5E model adherence, teacher–student role clarity, inquiry support, chemical accuracy, SSI integration, assessment quality, and language appropriateness.

Significant differences in model performance were observed. Claude produced the most comprehensive and pedagogically robust plans, with strong alignment to learning outcomes, effective SSI integration, and thorough assessments. ChatGPT offered structurally coherent plans aligned with the 5E model, but content depth was moderate. DeepSeek generated organized and practical plans, yet showed inconsistencies in 5E alignment and learning outcome coherence. Gemini and Microsoft Copilot performed weaker, with limited alignment to learning outcomes and more superficial chemistry content.

Overall, while all models generated broadly implementable lesson plans, their pedagogical quality varied significantly. The findings highlight the importance of model selection and prompt design in leveraging GenAI for chemistry education and suggest that RAP can be an effective strategy for enhancing instructional outputs.