What should we do when AI starts believing its own fiction?

As synthetic content feeds into training pipelines, the risk is no longer restricted to just hallucination and deepfakes but ...