Unlearning descriptive statistics