'Detoxified' language models might marginalize minorities, says study

A new study finds evidence that techniques used to 'detoxify' language models can amplify biases against certain minorities.