Humans already unable to tell the difference between generated and real. You get compounding of training on ever more fake data. You don’t see this as problematic?