Combining top-down moral principles with bottom-up reinforcement learning is a promising approach for developing safe and adaptable moral AI agents.