To be precise, it's actually a little bit the opposite; if something requires planning, then current AI will struggle it. It can draw on lots of context, but it is unable to plan anything based on that context. Creative writing, like any temporal art form, is all about planning and mentally "testing" different outcomes, which is why so-called "causal" generation is not very good at it.

Sure, "beam search" operates somewhat like planning, but each "beam" is still a causal generation, and thus doesn't represent anything like a concept to the system. Also, the evaluation of beams is still based on overall probability, in the context of causal prediction, and is thus driven largely by syntactic factors. Human writers don't literally "write" the concepts they're testing, and they don't evaluate the quality of different plot ideas based on the sentences they used to think them. So the human cognitive approach is arguably vastly superior to the causal LLM approach, when it comes to creative writing.