The big problem is that human values are similar to neural network weights that can't be clearly defined into true/false axioms like "human life is inherently valuable" and "murder is wrong". The easiest axiom to break is something like "every human life is equally valuable", or even "every human life is born equal".