Machine learning models, even when delivering state-of-the-art performance, can produce potentially offensive, inappropriate or toxic content and reinforce stereotypes or project social biases. Offensive behaviour can inadvertently be caused as part of any stage of the machine learning workflow, including data sourcing, preprocessing, and model training. The topic of offensive content is especially important in the field of Natural Language Processing (NLP). This session covers techniques and approaches for identifying, quantifying, and mitigating offensive content in language models (used in autocorrect, smart compose, and translation). The demo will feature a hands-on Python exercise with Azure Machine Learning where we leverage popular machine learning frameworks for detecting and mitigating offensive content in language modeling.