I should have been more clear, because I certainly agree with what you're saying here. However, I think we can pretty clearly define the target output of a mode

I should have been more clear, because I certainly agree with what you're saying here. However, I think we can pretty clearly define the target output of a model with whatever it was trained on; in this case, GPT is trained on text. You're right that the model doesn't "know" that necessarily, but it has been designed to produce convincing text when given a small amount of entropy to start with.