Ask HN: Should ethics be built into machine learning models?

US banks were recently sued by justice department for discriminating on race in approving loan applications. The bank argued that there was no field for race in the application making it impossible for the bank to discriminate on race. Justice department argued that they could tease out race from the applicants name and found systematic discrimination. As machine learning becomes more common, what steps if any should practitioners try to build into the models in order to prevent what is considered unfair discrimination? Even if excluding these fields, there are many other fields that can be used to derived that field, like names. How should practitioners approach these issues? As the models become more and more intricate and drift away from human comprehension, is it even possible to effectively conclude what factors are driving the model and the ethics of using those factors?