All existing LLMs (including those optimized for creative writing) are extremely bad at this, they tend to write in a narrow subset of American English sentence structure and idioms, even if you prompt them to imitate someone's style. This is inevitable due to English being prevalent in the dataset and RL murdering the variance.

AI slop reads unnatural even in English due to its lack of variance. And it heavily leaks into all other languages, even Ancient Greek.

RL absolutely murders variance. GPT-4o was an order of magnitude harder to prompt into sustained chain of thought than GPT-4, from day 1 in my experience.