Yeah true. Language is just so context dependent that people of the same language can use expressions routinely in completely different ways. (Although I have no doubts that with enough training data/complex enough models even this can be solved)