The problem is that we do not know "How" to engineer those principles. And that's what the entire field of AI alignment is working on. We know what we want the AI to do; the problem is we don't know how to make certain it does that. Because if we only get it 99% right then we're probably all dead in the end.