So I guess the question is, if the meaning of "correct" isn't defined, how did you arrive at code to do that thing? Are we talking neural networks here or what? Even those get tested, maybe not in an automated way, but you could at least make sure the results are better than your last iteration.
Take the facial recognition example. Give the whole thing a few images that it should fail on, and some it should pass. As you learn more about the system and where things fail or just don't add up, add tests there. Its likelier those parts of the code are bad too so refactoring isn't unheard of.
To a degree I completely agree with you that some things don't have a easily known answer in the real world. But adding tests to test known things like regressions, just seems like it should be a minimum of testing. I can't count the real world apps I've seen without a single "proof" that they don't regress state as things change. And they do, often and repeatedly. But dealing with external products and teams can be frustrating.
Take the facial recognition example. Give the whole thing a few images that it should fail on, and some it should pass. As you learn more about the system and where things fail or just don't add up, add tests there. Its likelier those parts of the code are bad too so refactoring isn't unheard of.
To a degree I completely agree with you that some things don't have a easily known answer in the real world. But adding tests to test known things like regressions, just seems like it should be a minimum of testing. I can't count the real world apps I've seen without a single "proof" that they don't regress state as things change. And they do, often and repeatedly. But dealing with external products and teams can be frustrating.