The linked article is arguing against a straw man that I don't think many ML engineers and researchers actually believe. Whether or not loss functions and mode

The linked article is arguing against a straw man that I don't think many ML engineers and researchers actually believe. Whether or not loss functions and model calibration are part of the "data" or the "algorithm" is just a question of semantics. Nobody who knows anything about ML would argue, as is incorrectly suggested by this article, that the choice of loss function plays no role in producing bias or inequity. The actual argument that most closely resembles the straw man in this article goes something like this. "The general project of using deep neural networks is extremely flexible and powerful. It is possible to build datasets and train deep neural networks so that the biases in those models are understood and acceptable. When existing models show bias, there are usually engineering solutions that can remove the bias within the paradigm of deep learning." Counter arguments to this reasonable claim are much more difficult to defend. The research community rewards "whistleblowing" much more strongly than finding solutions, which is generally a good thing. But a nasty side effect is that the majority of algorithm fairness research is pessimistic, despite huge improvements and widespread belief in industry that these problems are solvable using known techniques.