how to implement ai safely