By focusing on six characteristics, the study claims you could reach "near-perfect accuracy" at detecting AI deepfakes.