This paper proves that self-verification prevents model collapse in recursive training without relying on real data.