Generalizing topics and/or adding caveats about how something may not always be true is the easiest way for a LLM to increase the rate of providing factually accurate responses (aside from refusing to answer). Both of those strategies requires being more verbose.
Concise and to the point is not in the best interest of a LLM designed to give decent answers 99% of the time.
I also believe that OpenAI's RLHF process highly biased the model towards producing that sort of specific verbose, padded-out 'chatGPT speak' that we are seeing. The RLHF fine tuning process took outputs from the instruction-tuned model and A/B tested variations with human workers who may or may not have had the same feelings towards writing quality as many of us.
The process resulted in those verbose, interjection-laden responses that we see now, because that type of response was deemed 'better' (thumbs-up'd more) than the shorter, more-direct-but-less-impressive-sounding responses.
I use it for work so much I’d gladly pay way more than $20 just for access to GPT-4. It’s pretty terrible at programming but it still saves me loads of time generating the easy functions.
Anything remotely complex I still do by hand. But holy shit its nice having something do my boilerplate.