Harm reduction for NLP models