That are often either inaccessible or not accessible in a form that can be understood by humans within a human life span or not labeled enough so that even if there is bias, there’s no systemic way to detect such bias. The training phase of machine learning does add to the complication of explain-ability.