I have a suspicion that the reason a lot of LLM bots can output stuff akin to word salad is because a lot of their training data is word salad in the first place. If you're a competent reader pre-LLM, you can mostly avoid/discount word salad, so its prevalence doesn't really register.
My YOShInOn can't write text, at least not yet. (I am thinking about T5 to attach Mastodon tags) I think most generative models would be too polite to say something like that or would be more tentative.