That, as the article clearly lays out, is a separate but closely related subject. It's a concern founded on the same common misconceptions about AI that they ar

That, as the article clearly lays out, is a separate but closely related subject. It's a concern founded on the same common misconceptions about AI that they are talking about here. Machine learning models are not well equipped to handle the issues of prejudice or fairness, because those are, in the most literal sense, subjective. Machine learning models as a feature do not interact with subjects. Subjects are too computationally hard, so they were optimized out instead. As a latent result, any behavior implemented with machine learning models must be carefully curated, otherwise subjective patterns like prejudice and groupthink will pollute the model. This is a trivially predictable continuation of prejudice as a language feature: it's functionally much easier to generate a continuation out of a biased circular position than it is to generate one that includes objective consideration. In fact, objectivity itself is a feature that has not been implemented yet in any AI software. Prejudice, on the other hand, is so easy to implement that it happens by accident.