"You can imagine how if we wanted to create algorithms that could help inform judicial decisions or credit scoring systems or employment decisions and we did so by training those algorithms on existing, human-made decisions in these areas, then all of the biases underlying those human decisions woul