Training data
A great deal of the ethical problems associated with ML biases come from the training data of the models. If the underlying data is biased, so too will the models. In fact, there’s a good chance the models will learn a given feature that encodes the bias (gender, racial).