That said, and generalising from the success of formalisation of other previously informal domains (e.g formalisation of fair elections lead to insights like Arrow's impossibility theorem), I expect that formalising concepts of justice will lead to major progress.
that is, if you clean your training data properly and adjust for correlates of injustice, e.g. race. I could imagine two ways (I am sure there are more): 1) you only train it on data from one sub-population, e.g. whites / wealthy people / etc.. 2) you carefully correct wrong decisions and biases in your training data and don't even train on such factors as race.
and then, in applying the model, you don't specify those correlates, either.
this could be a small revolution.