“As we saw, machine learning algorithms optimizing solely for predictive accuracy may discriminate against racial or gender groups, while algorithms computing aggregate statistics or models from behavioral or medical data may leak compromising information about specific individuals.”