Large language model-based (LLM-based) assistants have become central to everyday tasks, yet despite consistent efforts to explain how these systems work, a transparency paradox persists: simply providing additional details does not necessarily lead to deeper user understanding or trust. This study presents findings from an artifact analysis of three popular LLM-based assistants, examining how these systems disclose knowledge boundaries, data provenance, and risk factors. Using a comprehensive evaluation framework across five task categories, we analyzed how each system managed transparency in different contexts. Our findings reveal significant variations in transparency implementation, with notable performance gaps in creative content generation and educational contexts. We identify key transparency paradox manifestations, including uncertainty maintenance failures, risk-correlated disclosure patterns, and critical thinking gaps. Based on our comparative analysis, we propose interaction design strategies to ensure transparency and user engagement on a more critical level, including contextual disclosure systems, reflective checkpoints, and domain-specific transparency adaptations.