> Does the chatbot imagined something in their head before describing it to me as a prompt?

You're attributing to the model what is the merit of the training data. It's culture that knows what would be a great answer to your task, of course, when culture is loaded up into a brain or an AI.

What I mean is that it doesn't matter the substrate - as long as it learned the distribution, then it can imagine coherent and amazing things. That's all the merit of the training data though. The brain or the model just dutifully carry that in a compact form that can be unfolded in new ways on demand.