It contributes to the impression of dodginess that most of the time it's "LLMs are totally inscrutable, we can't tell anything about the relationship between training data and output, we can't explain any decisions" and then when it becomes inconvenient they're "nope, definitely didn't happen"