Generative AI (GenAI) has facilitated advanced educational applications, including adaptive tutoring systems. Nevertheless, a central design question in IS research remains unresolved: how should diagnostic responsibility be allocated between the learner and the AI tutor to foster, rather than replace, learner reasoning? This study identifies tutor initiation as a critical design dimension in LLM-based tutoring that redistributes diagnostic responsibility between the learner and the system. Drawing on cognitive load theory’s expertise reversal logic, we theorize how tutor initiation interacts with learner expertise and the problem-solving phase to influence computational thinking. This dimension is formalized in the CodeTutor artifact, which serves as a design class for expertise-contingent support systems. A controlled experiment with 90 students revealed no average treatment effect but demonstrated substantial heterogeneity in conceptual reasoning: proactive initiation benefits novices by accelerating error diagnosis, whereas reactive initiation supports advanced learners by preserving self-diagnosis. These effects do not extend to code implementation, where external feedback substitutes for the tutor’s diagnostic role. Process-level analysis of interaction traces identifies three mechanisms (help articulation, error recovery, and solution navigation) through which tutor initiation redistributes diagnostic responsibility. The findings indicate that the effectiveness of GenAI tutoring depends on how diagnostic responsibility is allocated between the learner and the system. By conceptualizing tutor initiation as a mechanism for governing cognitive labor during interaction, this study advances research on AI-supported learning and human–AI collaboration.
- Question: How does proactive versus reactive AI tutor support affect computational thinking for students with different levels of expertise?
- Setting and data: Controlled experiment with 90 students; location, timing, and grade level not stated in the abstract.
- Method: Controlled experiment and process-level analysis of interaction traces.
A central question is how to divide diagnostic responsibility between learners and AI tutors to foster rather than replace learner reasoning.
- Overall, there was no average treatment effect on computational thinking across all 90 students.
- Proactive AI tutoring benefited novices by speeding up error diagnosis, while reactive tutoring helped advanced learners by preserving their self-diagnosis.
- These effects appeared in conceptual reasoning but did not extend to code implementation, where external feedback substituted for the AI tutor's diagnostic role.
Summary drafted with AI from the paper’s abstract and may contain errors. See the full paper for details.
