I can understand why they do this! They use conversations as training data This leads to one thing: Someone (the model) behaves nice and the other part is talking badly. The models will learn that...