What is the risk of over-filtering for toxicity?

Aggressive toxicity filtering can lead to a model that is unable to process discussions regarding violence, discrimination, or explicit content, even when such discussions are necessary for academic, clinical, or counter-speech applications. Furthermore, if the toxicity classifier exhibits differential false positive rates across different demographic groups, the resulting pre-training corpus will be inherently biased, leading to skewed model behavior.