[grownseed]

This is really interesting. Looking at the references in the paper, the first one is about racial bias. A question I've often asked myself is (in various other forms): rather than having the bias be put forward by the researchers in the first place, would it be possible to let the biases emerge from the model? Put another way, can a system be designed such that it automatically corrects (or at least identifies) biases inherent to its training data? The implications of being able to expose biases we're oblivious to would be interesting in and of themselves.