Large language models (LLMs) are making waves in education, promising personalized tutoring and automated grading. But how do they handle inconsistencies between their internal knowledge and the information found in textbooks? Researchers have developed a new dataset called EDUKDQA (Educational Knowledge Discrepancy Question Answering) to test the robustness of retrieval-augmented generation (RAG) systems—which combine LLMs with information retrieval—when faced with conflicting information. EDUKDQA presents multiple-choice questions across science and humanities subjects, introducing subtle knowledge discrepancies to mimic real-world scenarios where textbook information might differ from an LLM’s training data. The results? While LLMs excel at multi-hop reasoning and handling distant context, they stumble when integrating their own knowledge with facts from the provided text. This challenge is amplified when the questions involve knowledge discrepancies, leading to a significant performance drop of 22-27% in RAG systems. Surprisingly, traditional keyword-based retrieval methods like BM25 often outperform more complex approaches in this specific domain due to their ability to pinpoint academic terms. This research highlights the need for more sophisticated methods to reconcile conflicting information, particularly as LLMs become increasingly integrated into educational settings. The future of AI in K-12 hinges on addressing this challenge to ensure accurate and reliable learning experiences.