An outsider can throw test data points at the model and see how it behaves. For example, whether changing name, race, sex, or variables correlating with such critical variables, alter a decision. Pointing out these flaws could then lead to better systems, or not using machine learning on some case where input data is unsuitable.