I don't think you really understand the kinds of details that AI picks up on. If that data pool shows that people who apply for loans on a Tuesday and have a sa

I don't think you really understand the kinds of details that AI picks up on. If that data pool shows that people who apply for loans on a Tuesday and have a salary whose third digit is 7 almost always default on loans, then neural networks will pick up on that, regardless of the connection to reality. An algorithm, especially a trained algorithm, should be assumed biased until proven otherwise. It's not humanly possible to cleanse and normalize a dataset to a level that will prevent all possible overfitting ever, and even if you have godlike precision in squashing irrelevant data, it still depends on your personal definition of unbiased. As someone who has been boots on the ground for political surveys, I'd frequently asked surveys that respondents from both parties would comment were biased against their own side. I don't trust human judgement, and I don't trust the judgement of human tools.