Think of the million+ more "invisible" things where this is happening across models as well and you'll see why "synthetic training data" is exactly what people predicted it would be (garbage feeding on garbage feeding on garbage to produce more and more incoherent garbage).