Retrieval-augmented generation (RAG) is widely used to enhance AI’s knowledge with new or changing information, but real-world complexities challenge its reliability. This talk explores how Large Language Models (LLMs) process retrieved evidence, why they often fail to utilize it correctly, and how misinformation risks emerge from unreliable or insufficient context. Based on recent research, I will highlight key pitfalls in LLM-driven fact verification and discuss strategies for improving fact checking robustness.