You might want to look up how it works then, because it truly is just glorified autocomplete. It can appear to output some seriously cool things, but

1. Remember the output is based on what the model was trained on. If its output is good, it’s only because the model was trained on a shitload of examples to produce a well mapped token graph. Examples like every single thing they can scrape off the internet damn the copyrights…
2. It’s only ever going to be an illusion of intelligence based off of the associations humans have already given to words and other tokenized things. LLMs will never grow past their training data.
3. While they can have the illusion of intelligence, they only ever are associating tokens. Sure, they can have large sets of input tokens, too, to relate the output tighter to what you want, but it’s all just mathematical associations!

The output works out a lot of the time when using models and GPUs way bigger than would fit on most anyone’s home computer… but it’s still just associations of tokens. They don’t know that all those tokens paint a picture that means token #3745, the protagonist, has ongoing motivation to interact with token #3758, the love interest. If your input instructions don’t directly associate those two with sheer reppitition, the "AI"s would basically all just wander all over the place as far as their relationship was concerned.

Magnify that kind of “vapid story” type problems across all aspects of a story that aren’t basically pre-written in the prompt anyways, and it turns out “AI” is actually extremely shit at anything requiring actual intuition, understanding, and basic intelligence no matter how much electricity is thrown at it!