Efficiency gains can paradoxically worsen outcomes, termed the strong version of Goodhart's law, seen in education, politics, and machine learning's overfitting. Mitigations include better proxy alignment, regularization, noise injection, early stopping, and adjusting system capabilities.