Misaligned by Design: Incentive Failures in Machine Learning