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.
