3 ms·

> I'm saying no policy could ever be defined well enough to cover all possible contingencies of a messy, human world. "All models are wrong, but some are usefu

• version_five 5y ago
  Not when people are punished as a result of them. It's fine to make a generalization that works on average for some low stakes, benign decision, but when it materially affects real people, it is not right to generalize. This is the core argument imo against algorithmic decision making of this kind.
  • julienb_sea 5y ago
    You can make the exact same argument in reverse. When decisions materially affect real people, it is unacceptable and fundamentally unfair to rely on human judgement, along with its biases and capriciousness. Clear, transparent policy applied evenly and without bias is a more fair approach for people.
    • version_five 5y ago
      No, it doesnt work that way. By projecting real, complex decisions down to some low dimensional set of criteria, you automatically throw away information and generalize. It's only fair in the sense that the criteria is the same, but it's necessarily based on some kind of average outcome and so effectively sacrifices outliers. Its is still extremely unfair to inputs (who are actual people in the cases we mean) that exhibit characteristics that reliably fit the prediction on average, but that don't conform to the prediction.