from problem formulation to data collection, from labeling to model design, from training to deployment. There is no single point where "the bias happens," which means there is no single point where it can be fixed.