LLM stands for “large language model”. It’s producing strings of text that at statistically similar to sequences in the training data. External reality isn’t relevant. It’s at least possible to run the result through a confidence filter, but that would be a feature of the tooling, not the model.