The problems with LLM's are twofold:
Firstly, they only have limited space for context, so they tend to "forget" the context beyond the immediate and the overall story, leading to many inconsistencies. Of course, you can "remind" them, but it ruins the smooth flow, the spontaneity and the immersion in the moment. This may improve with technological advances, but I believe it will still be a limiting factor.
Secondly, LLMs deals with words (or tokens), not with concepts. They generate responses to prompts that are deemed to be plausible according to the material they were trained on, but every now and then you can see the discrepancies.
For example, an LLM can write something like this:
As Jack watered the beanstalk, it grew, reaching up to the clouds, over 20 meters tall.
But they can still be fun to experiment with.
This would eradicate Ghosting. But also remove potential connections you may make with a partner. Would be willing to try it, but I'm already worried of a I Have No Mouth and I Must Scream scenario with all this AI.
AI's have the capacity to manipulate words, not bodies and minds.