Large language models (LLMs) are impressive, but they can also be confidently wrong. Knowing when to trust an LLM is a huge challenge, especially for critical applications. Existing methods for gauging AI certainty often fall short. They either need access to the model's inner workings, which isn't always possible, or rely on superficial text comparisons that miss subtle differences in meaning. Researchers have been exploring ways to cluster similar answers to a question. The idea is that a wider spread of different answers implies more uncertainty. However, simply counting the number of different answer clusters isn’t enough. A new study introduces a clever technique called Contrastive Semantic Similarity (CSS) that digs deeper into the meaning of AI-generated text. Inspired by a powerful image-text model called CLIP, CSS extracts "similarity features" between different answers. These features capture the nuanced relationships between texts, going beyond surface-level comparisons. By applying CSS to group similar answers, the researchers found they could better estimate the LLM’s uncertainty. This method could be particularly useful for "selective answering," where an AI only gives an answer when it's confident enough. Experiments on question-answering datasets showed that CSS outperformed existing methods, correctly identifying and rejecting unreliable answers more often. This research is a step forward in building more trustworthy and reliable AI systems. By figuring out when AI is unsure, we can make better decisions about whether to rely on its output. The research team plans to explore how to further refine this approach and apply it to broader language tasks.