Open Access

The rapid advancement of large language models (LLMs) has introduced powerful artificial intelligence (AI) tools into educational environments. While AI assistants offer potential benefits for learning, concerns about over-reliance, reduced critical thinking, and impaired skill development have emerged. This thesis investigates how the timing of AI support (Just-in-Time vs. Always-On) and reflective mandates (Rationale-Required vs. Rationale-Optional) influence creative performance, learner autonomy, and critical engagement in AI-assisted data-science problem framing. Through a controlled 2x2 within-subjects factorial experimental design with 66 postgraduate participants, the study examines expert-rated idea quality, semantic diversity, perceived agency, AI reliance, cognitive load, and reflective reasoning across four AI-assisted conditions. The results show that Just-in-Time support and required reflection are independently associated with higher idea quality, greater agency, lower AI dependence, and more selective engagement with AI suggestions. The study does not assess delayed or long-term learning transfer; future work with longitudinal designs is needed to determine whether the immediate benefits observed here translate into durable skill development.