I've found that these patterns are very specific to OpenAI models. My guess is that they're being introduced through some kind of post-training fine tuning process. It's not something inherent to language models, and other models are typically far more natural and variable in output structure.

Small discussion when first posted (15 points, 19 days ago, 5 comments) https://news.ycombinator.com/item?id=46133941

Archive link: https://archive.ph/IX9bP