The rapid proliferation of misinformation on social media underscores the urgent need for scalable digital-literacy instruction. This study presents the design and evaluation of a Generative AI-enhanced serious game system that integrates Large Language Models (LLMs) to drive adaptive non-player characters (NPCs). Unlike traditional scripted interactions, the system employs role-based prompt engineering to align real-time AI dialogue with the Currency, Relevance, Authority, Accuracy, and Purpose (CRAAP) framework, enabling dynamic scaffolding and authentic misinformation scenarios. A mixed-method experiment with 60 undergraduate students compared this AI-driven approach to traditional instruction using a 40-item digital-literacy pre/post test, the Intrinsic Motivation Inventory (IMI), and open-ended reflections. Results indicated that while both groups improved significantly, the game-based group achieved larger gains in credibility-evaluation performance and reported higher perceived competence, interest, and effort. Qualitative analysis highlighted the HCI trade-off between the high pedagogical value of adaptive AI guidance and technical constraints such as system latency. The findings demonstrate that Generative AI can be effectively operationalized as a dynamic interface layer in serious games to strengthen critical reasoning. This study provides practical guidelines for architecting AI-NPC interactions and advances the theoretical understanding of AI-supported educational informatics.