Artificial intelligence: how little has to go wrong?