A major AI training data set contains millions of examples of personal data