It breaks simple random sampling meaning that you're optimizing purely for some unknown subset. That becomes a problem when someone someone realizes that Netflix is spending millions of dollars to have a recommendation engine that ends up being racist.