The first goal is to align the model to the least common denominator of ethics. Even if there is diversity in values, everyone agrees that murder is bad, or that you shouldn't torture your neighbor for fun.