How matters more than whether: generative AI role paradigms and learning in K–12 education

In K–12 education, how should students interact with generative AI to maximize learning while minimizing risks? This study compared three AI role paradigms, namely AI-Directed, AI-Supported, and AI-Empowered, against a traditional control in a 4-week quasi-experiment with 176 Grade 8 students in China. Outcome measures spanned epistemic, practice, and affective domains, supplemented by interaction logs and risk indicators. While main effects on learning outcomes were not significant, students in the AI-Empowered and AI-Supported conditions showed greater gains in deep learning approaches and higher-quality AI interactions than those in the AI-Directed and Control conditions. In contrast, the AI-Directed paradigm was associated with elevated risks, including higher textual similarity to AI outputs and increased self-reported academic misconduct. Mediation pathways and cognitive load differences were not supported. These findings suggest that how learners are positioned relative to generative AI may matter more than whether they use it, as active paradigms appear to foster deeper engagement without increasing risks, whereas passive reliance may carry hidden pedagogical costs.