Some kind of other thing. It is output based on its inputted training data, so pretty much like a fancy version of the autocomplete on a mobile phone (before that became LLM-led).

If a topic is widely factual in the training data, the output is more likely to be factual.