Provided the data isn't wrong, we can typically put tight statistical error bounds on the output of various ML algorithms. Most people's understanding of how algorithms can be "biased" is completely wrong. There was a whole furor a while back about algorithms discriminating due to language differerences in the data for people of different races. Guess what; if you give race as a parameter to any half-decent algorithms trained on that data, the algorithm will learn that the written data contains those biases. Decent algorithms trained on true data are pretty much guaranteed to have results within a very very small margin of reality.

> Guess what; if you give race as a parameter to any half-decent algorithms trained on that data, the algorithm will learn that the written data contains those biases.

You know what? Even if you do not give race as a parameter; an algorithm could be biased. It can easily learn race from secondary or tertiary parameters.

I don't think you understood my post; giving race as a parameter allows ML algorithms to detect and counteract human racism.

If an ML algorithm notices a disparity along e.g. racial lines, it's because it's actually there, not because a human imagined it.