Google DeepMind researchers propose a new way to test whether AI chatbots actually understand morality or just mimic it, moving beyond current surface-level evaluations.