Large language models (LLMs) are impressive feats of artificial intelligence, capable of generating human-like text that can be indistinguishable from something a person might write. However, these powerful tools have a hidden flaw: they can sometimes generate incorrect or entirely fabricated information, a phenomenon known as 'hallucination.' This poses a significant challenge, especially in fields where accuracy is paramount, such as medicine, law, and finance. How can we tell if an LLM is hallucinating? Researchers are actively working on this problem, exploring various methods to detect these AI-generated falsehoods. A new study proposes a supervised learning approach using a surprisingly small set of numerical features derived from token probabilities. These probabilities, generated by other LLMs acting as 'evaluators,' provide clues about the likelihood of a given text being a hallucination. The research team trained two simple classifiers—a logistic regression model and a simple neural network—using just four features. These features capture the minimum and average token probabilities, the maximum probability deviation, and the minimum probability spread. The results are promising, showing that this approach can effectively detect hallucinations in various tasks, including summarization, question answering, and knowledge-grounded dialogue. Interestingly, the research also revealed that using different LLMs as evaluators can improve detection accuracy, suggesting that diverse perspectives in AI can help identify and mitigate biases. While this approach shows great potential, it's not without its limitations. The researchers acknowledge that the method's effectiveness varies across different datasets and tasks. For instance, it struggles with more nuanced forms of hallucination where the falsehoods are less obvious. Furthermore, the method currently relies on binary classification (hallucination or not), which doesn't capture the varying degrees of severity that hallucinations can exhibit. Despite these limitations, this research offers a valuable new tool for detecting hallucinations in LLM-generated text. The simplicity and efficiency of the approach make it a promising direction for future research, paving the way for more reliable and trustworthy AI systems.