They can and do, but they want the training models to come from highly moderated sources otherwise everywhere AI chatbot would be spewing the most racist parts of 4chan because people would train it that way as a joke.