The problem here is that modern machine learning is rarely more than a computerized way of building statistical models based on observed data. And the bias mitigation techniques I've read about all revolve around (indirectly) manipulating those models based on what results they produce. This can lead to various problems in the long run.