Methods, systems, and computer programs for identifying offensive content. A method can include for each particular responsive message received in response to an initial message: providing the particular responsive message as an input to a machine learning model trained to predict a likelihood that an initial message includes offensive content based on processing of a responsive message received responsive to the initial message, processing the content of the particular responsive message through the machine learning model to generate output data indicating a likelihood that the initial message includes offensive content, and storing the generated output data. The method can further include determining, based on the stored output date for each of the responsive messages, whether the initial message likely includes offensive content, and based on a determination that the output data for each of the responsive messages indicates that the initial message likely includes offensive content, performing a remedial operation.