DeepSeek-assisted physics instructional design: An empirical study of high school physics teaching

Generative artificial intelligence (GenAI) has shown growing potential in supporting teacher education, yet little is known about how preservice teachers collaborate with AI during instructional design (ID). This study examined the effects of DeepSeek, a Chinese large language model, on preservice physics teachers’ ID performance and cognitive processes. A quasiexperimental design was conducted with 220 participants completing a 45-min ID on Newton’s third law, either with or without DeepSeek assistance. Data from IDs, interaction logs, and interviews were analyzed through statistical tests, epistemic network analysis (ENA), and thematic coding. Results indicated that the DeepSeek-assisted group outperformed the control group in analytical and structural aspects of ID, such as teaching content analysis and learner analysis. ENA revealed that high-performing participants exhibited proactive “problem representation–AI interaction–monitoring” patterns, while low-performing participants relied on reactive prompting. Interviews further confirmed that AI was helpful in contextual introductions and experiment design but also revealed issues such as outdated curriculum references and overly absolute expressions that lacked sensitivity to teaching contexts. The findings suggest that GenAI can serve as a cognitive scaffold when used strategically and highlight the need to integrate AI literacy and metacognitive training into teacher education.