How do you know you have the correct model, and isn't making the system more racist instead of less?

Machine learning is also very opaque.