Large language models (LLMs) sometimes produce “hallucinations”—outputs that are factually incorrect or contradict previous responses. But what happens when an LLM hallucinates even though it *actually* possesses the correct information? New research explores this intriguing phenomenon, distinguishing between two types of hallucinations: those stemming from a lack of knowledge (HK[-]), and those occurring *despite* having the knowledge (HK[+]). Think of it like this: sometimes, the AI genuinely doesn't know the answer. Other times, it's like it knows the answer but gets confused and says something else entirely. Researchers developed a method called 'Wrong Answer despite having Correct Knowledge' (WACK) to create model-specific datasets of these HK[+] hallucinations. They found that by analyzing the AI's internal states, these two types of hallucinations look different. This means that one day, we might be able to predict when an AI is about to hallucinate, even *before* it gives a wrong answer! The study also revealed that different AI models hallucinate in unique ways, suggesting that personalized approaches to hallucination detection and mitigation will be crucial. This is a big step towards understanding why AI sometimes says things that don't make sense, and it opens exciting possibilities for building more reliable and trustworthy AI systems in the future.