A more familiar example might be fuzz testing. You could fuzz a program by just passing the fuzzer totally random data to mutate, and it's true that this might find a crash that wouldn't be found any other way. However, most of the time you will have much better results, and achieve them much more quickly, if you pass the fuzzer something that looks like an actual expected input to the program.

Zero-knowledge ML is still interesting if your goal is to indisputably demonstrate the act of learning, just not if your goal is to make the smartest program in the least time.