In the most general sense of the word bias, there can be biases that arise purely as a result of model architecture. For example, I was messing around with GANs

In the most general sense of the word bias, there can be biases that arise purely as a result of model architecture. For example, I was messing around with GANs about a year ago, and kept on getting checkerboard-like patterns in the generated images. This was a result of how I had set up the convolution layers: It was a bias introduced by model architecture. In terms of the "unfairness based on what group you belong to" kind of bias, that's almost always a result of bad input; you'd have to work really hard to make an inherently racist architecture. So the problem is introduced in the training data. But it could be solved either by getting better training data, or by making clever improvements to your algorithm to work around the limitations of the training data. Often changing the algorithm is easier than improving the data, which I think is the point this article is making. As an example of such a change, let's say we're designing an algorithm to estimate credit score, and we're worried about unfairness based on group membership. Let M be a variable representing what groups a person belongs to: race sex, etc, and let S be the output of the algorithm: that person's credit score. An easy thing to do would be to not allow M as an input to our algorithm. That way we can say that the output "does not depend on group membership". The problem is that we may still have proxies for M as inputs to the algorithm, and if the dataset is biased, that algorithm will still learn to use those proxies to discriminate based on group membership. A better method is surprisingly to keep M as an input during training of the model. Then, at inference time, average over M to get the final output of the model. That still erases the information about what group a particular person belongs to, but it also has the advantage that the model isn't incentivized to learn proxies for M during training, since it has access to M directly.