Trusting the output of an abstraction explicitly designed, engineered, and tested by other humans with an actual understanding of the layer beneath you is not comparable to trusting a non-deterministic, statistical next-word-predictor with no mind or ability to be morally responsible for what it produces.