“We should always be suspicious when machine-learning systems are described as free from bias if it's been trained on human-generated data. Our biases are built into that training data.”