Caution about automation is quietly conservative

I take the fairness concerns seriously, but the counterfactual is never a fair human system. It is the status quo, which is arbitrary, unaccountable and invisible. Human loan officers were biased for a century and nobody could audit them. At least a model can be measured, corrected and measured again. Demanding perfect interpretability before deployment functionally means keeping systems we already know are worse, and the people who pay for that delay are the ones currently being refused by someone having a bad afternoon.