The issue I take with the concerns raised in the paper regarding:<p>"A language model that has been trained on such data will pick up these kinds of problematic associations."<p>"Such data" is basically everyday, normal discourse, and some of the "problematic associations" are training that includes phrases like "woman doctor", "both genders", etc. While I get the point, this itself is a biased interpretation of discourse and would be worrisome imho to have people filtering models with their own biases vs the language as it's used by the vast majority of people.