One possibility is that alignment isn’t just about cultural distance in values, but also about textual representation in the training data