The article discusses the issue of hallucination in language models, where they generate confident-sounding responses without actual measurement or certainty. When asked for a forecast and confidence level, a language model may respond with a high confidence level, such as "90% confident", but this is simply a word choice rather than a genuine measurement. The model's actual uncertainty comes from two sources: epistemic uncertainty, which decreases with more data, and aleatoric uncertainty, which is inherent to unpredictable events and never goes away. This highlights the need for a more nuanced understanding of language model outputs and the importance of distinguishing between genuine confidence and generated text.