Large language models (LLMs) like ChatGPT are impressive, but they sometimes "hallucinate," meaning they confidently generate incorrect information. This poses a major challenge for real-world applications where accuracy is crucial. New research introduces a clever method called "semantic entropy probes" (SEPs) to detect these hallucinations. Traditional methods for identifying AI hallucinations rely on generating multiple answers to the same question and seeing how much they vary. If the answers differ wildly, the AI is likely uncertain and hallucinating. This approach, while effective, is computationally expensive. SEPs offer a much cheaper solution. They work by examining the AI's internal state after processing a question, but *before* generating an answer. Researchers found that this internal state already contains information about how certain the AI is about the answer. Like a poker player's tell, the AI's hidden thoughts reveal its confidence level. This is done by training a simple linear model, a "probe," to detect semantic entropy, a measure of uncertainty, directly from the AI's hidden state. The results are promising: SEPs can predict hallucinations with surprising accuracy and are far less computationally intensive than previous methods. Moreover, these probes can be trained without needing labeled data or examples of correct answers, unlike previous methods. This makes them easier to train and improves generalization. The research suggests that LLMs possess a kind of internal uncertainty meter, even before they articulate an answer. This discovery opens up new possibilities for making AI more reliable and trustworthy. While SEPs are not as accurate as the most expensive methods, their low cost and ease of use make them a significant step towards catching AI hallucinations in the act. This work also hints at deeper insights into how LLMs represent knowledge and uncertainty, paving the way for more robust and transparent AI systems in the future.