But randomness are not enough, it's possible the first generated version correlate with your preference and that situation does not really matter why are you stopped generating new password.
Predictability of this scheme very good, because the strict rules. All implementation using common english words. For words smaller length, easy spelling, less ambiguity are preferred. Mixing nouns verbs and adjectives are more meaningful, order also adds more meaning. Word count very limited. Preference on the separator character also very rigid. Personal preference more known, because hidding your preference on all english words very hard and mostly unconscious.
Eg the scheme doesn't do its job well if you don't know the word, so the dictionary can be reduced by that much
Random numbers picked by humans are notoriously biased. I'm guessing it's even worse when you ask them to come up with random words.