This is why AI is complete snake oil. They're trying to encode a recognizer, but they can't even define what it is they are actually feature extracting. Only that it's "typical ML techniques" that the rest of the industry doesn't want to call bullshit on because no one wants to give someone the idea that they shouldn't be allowed to use something if they can't explicitly articulate the function.

There's some things where screw it, why not. This is not one of them.