My understanding is that it isn’t the algorithm that’s biased per se, it’s the training data. If, for example, in the United States a black person were more lik

My understanding is that it isn’t the algorithm that’s biased per se, it’s the training data. If, for example, in the United States a black person were more likely to be investigated and prosecuted than a person of another race, then the crime statistics would be biased. If these statistics were then fed into a machine learning model, that model would actually be predicting an individual’s likelihood of conviction within the current system, rather than the likelihood of their actually committing a crime. I’m not even making the claim here that this is the case, but it is a reasonable example. Garbage in, garbage out.