Failures of Fairness in Automation Require a Deeper Understanding of Human-ML Augmentation