Data-driven algorithms are discriminating based upon undesirable/illegal vectors; they are utterly amoral in optimizing their solutions. Even if the algorithm does not have access to the "Age" field, then there are plenty of proxies, like what reunion tour you liked. And the same goes for race, gender, sexual identification, religion, etc.
To solve this we either need the training data to have no illegal/undesired discrimination, or we make the system moral. I think the first is impossible, and the second is what we will do sooner or later.
To solve this we either need the training data to have no illegal/undesired discrimination, or we make the system moral. I think the first is impossible, and the second is what we will do sooner or later.