It is a bold assumption to think that discord hasn't always been using harvested text/voice/video data to train models.