Because the computer in this case is relying on a vast repository of human knowledge encoded in these languages to go from 1 to 3. Machine code was probably not in the training set.

Would be interesting if you could train a GPT-N model to be a compiler somehow.