Large language models (LLMs) are impressive, but they can be frustratingly overconfident, even when spouting incorrect information. Imagine an AI confidently telling you that the sky is green – annoying, right? New research introduces a clever technique called LACIE (Listener-Aware Calibration for Implicit and Explicit confidence) to address this overconfidence issue. Instead of just training LLMs to answer questions correctly, LACIE trains them to express their confidence in a way that makes sense to a listener. It's like teaching an AI to say "I'm not sure" when it's actually not sure. Researchers simulated a 'listener' AI that judged the speaker AI's answers based on how confident they sounded. This feedback helped the speaker AI learn to express uncertainty appropriately. The results are promising: LACIE-trained models were much better at signaling uncertainty when they were wrong, leading to a 47% reduction in incorrect answers being accepted by human listeners in a study. This approach also generalizes to other datasets, improving truthfulness on questions designed to test common misconceptions. While there's still work to be done (sometimes the AI's confidence comes across as a bit rude!), LACIE represents a significant step towards making AI more trustworthy and reliable. It's all about teaching AI not just what to say, but how to say it – with the right level of confidence.