This depends a lot on how the model was fine-tuned. There are many LLaMA tunes that are not at all formal, for example.

With ChatGPT specifically, some forceful prompting can make it sound much better.