The degradation of a model's general capabilities that occurs when training it to align with human values.