Its not model collapse nor does it have anything to do with training data frequency. It's simply RLHF where the humans hired to tune the conversational style of these LLMs preferred certain idioms over others and so the reward function for these LLMs gravitated toward using them.

If LLMs generated text based on training data frequency they'd likely be some of the most vulgar and hostile things ever created. The internet is full of insults, profanity, and low effort content. The repeated phrases are a side effect of reward optimization rather than some kind of model collapse.