Large pretrained language models, trained on vast amounts of online data, often acquire toxic behavior and biases from the internet. This is a concern because these models are powerful and widely used in natural language processing tasks. To safely deploy them, safety controls over the model generation process are needed. --- Why it matters: This matters to AI researchers because it highlights the need for more robust methods to mitigate the negative effects of toxic behavior in language models, which can have significant consequences on user trust and model performance. Source: https://lilianweng.github.io/posts/2021-03-21-lm-toxicity/