The answer is not better detection software. It's redesigning assessment around things AI cannot fake on someone's behalf.