The difference between real human imagination and this is that when we imagine, we try to focus on our experiences and find a safe place. What this does is focus on what others have said when they were asked to imagine in the training input it received.

You have given a very good example of why these LLMs are not conscious. Being able to have thoughts independent of prompt is where true AI begins. If we ever get such LLMs then we are close to AGI.

Such LLMs are hard because language/speech is not naturally occurring without humans. So multimodality is essential as well. An AI that functions only on your command is a machine, an AI that has motives/goals beyond orders is a sentient being.