Large language models (LLMs) are impressive, but they have a tendency to 'hallucinate' – generating text that sounds plausible but is factually incorrect or completely made up. This isn't just about LLMs getting facts wrong; it's about understanding why they sometimes fabricate information, especially when used with retrieval augmented generation (RAG), where they pull from external documents to answer questions. Researchers are tackling this challenge head-on, developing new ways to catch these hallucinations in the act. A recent study introduces a novel approach using a dataset called HalluRAG, designed to detect these 'closed-domain' hallucinations, where an LLM invents information even when it has access to the correct answer in provided documents. The key innovation is focusing on recent information, added to Wikipedia *after* the LLM's training data cut-off, ensuring the AI hasn't seen it before. This lets researchers precisely control what the LLM knows and see when it starts making things up. The researchers built a classifier using the LLM's internal workings to predict whether a generated sentence is a hallucination. Results showed promising accuracy, particularly with the Mistral-7B model. Interestingly, the type of question—whether it had a definitive answer within the provided context or not—played a significant role. Hallucinations for answerable and unanswerable questions seemed to be encoded differently within the LLM, leading to much higher accuracy when the classifier was trained separately for each type. This suggests that LLMs process and represent these two types of questions differently. While HalluRAG showed limitations in generalizing to other datasets, it highlights the need for diverse and carefully crafted training data to build robust hallucination detectors. The next step? Creating even richer datasets and refining these classifiers to make LLMs more reliable and trustworthy, paving the way for their wider adoption in critical applications.