We often talk about AI being "confident" in its answers, but what does that even mean? A new research paper, "On the attribution of confidence to large language models," dives deep into this very question, exploring whether Large Language Models (LLMs) actually possess something akin to human confidence. It's a fascinating philosophical puzzle. We tend to think of confidence as a feeling, a subjective experience. But can algorithms have feelings? The paper argues that when scientists talk about LLM confidence, they aren't just using a convenient metaphor. They're making a real claim about the internal state of these AIs. This raises a thorny metaphysical question: can an LLM, a complex mathematical function trained on text data, actually *have* confidence? While the evidence is still out, the researchers suggest it's plausible, pointing towards how internal LLM representations of truth and falsehood might be interpreted as belief-like states. Even if we grant LLMs this ability, another puzzle emerges. How do we measure AI confidence? The research identifies the existing techniques – methods like prompting LLMs to report their certainty, observing the consistency of their answers, and analyzing the probabilities they assign to different outputs. But, the paper reveals serious flaws. The stochastic nature of LLMs, combined with technical details about how they generate responses, can distort these measurements, leading to questionable conclusions. The challenge lies in finding a clear link between the mathematical outputs of an LLM and the concept of subjective certainty. Output probabilities are tied to words, while beliefs are about the meaning of words—and that bridge isn't so easy to build. Even if LLMs can, in principle, form confident beliefs, current methods may not accurately capture that aspect of their operation. This is a critical insight as we increasingly rely on AI in decision-making. If we can't tell whether an AI is truly confident, how can we trust its judgments? Further research into how LLMs represent knowledge and form "beliefs" is crucial. As AI continues to develop, understanding the nature of its certainty is not just a philosophical exercise; it's a necessity for building trustworthy and reliable AI systems.