Seemingly trivial differences in training data can skew the judgments of AI programs—and that’s not the only problem with automated decision-making