I think my caveat of domain knowledge side steps this issue you bring up. It's the entire reason engineering schools teach the "long way" when you're going to r

I think my caveat of domain knowledge side steps this issue you bring up. It's the entire reason engineering schools teach the "long way" when you're going to rely on the "short way" 99.99% of the time. Also the same could be true for digital calculators when first created. The neural nets (aka humans) that were initially trained to design the system can take short cuts for example on improperly returning or applying functions such as square roots ect. It still confidently gives you the incorrect answer. Given sufficient training data (IE the human) the system improves to return better more accurate results.