LLMs got all these patterns from humans in the first place*. They're common in LLM output because they're common in human output. Therefore this argument isn't

LLMs got all these patterns from humans in the first place*. They're common in LLM output because they're common in human output. Therefore this argument isn't very reliable. If P is the probability that a text containing these patterns was generated by an LLM, then yes, P > 0, but readers who are (understandably) tired of generated comments are overestimating P. * Edit: I see now that the GP comment already said this.