Transparency of a robot's actions and intentions is important for trustful human-robot collaboration. Current approaches for creating transparency through explanations mostly follow the approach "more information creates more transparency", assuming cognitive load remains manageable. However, this ignores human active reasoning, which can, if not too demanding, create a better understanding from less information, as found in education literature. To explore this self-explanation effect, we compared three explanation structures (fully-specified, under-specified, no explanation) that induced different levels of active reasoning demand (low, medium, high). In a controlled laboratory study, 36 participants observed a robot completing rule-based classification tasks of varying difficulty (easy, moderate, hard) across all explanation conditions. We found that a moderate reasoning demand, elicited through under-specified explanations, produced the best understanding of the robot's actions compared to full or no explanations. This challenges the current "more helps more" approach and may help design more effective explanations for creating transparency.