Nor is the LLM output tightly coupled to the training data. They'll "eagerly"[1] fill in the blanks wherever it sounds good.

[1] here I don't mean to imply agency, just vigor.