When Models Eat Their Own Output, Lineage Is the Only Defence

Synthetic data is now training the next generation of models. Without a chain of custody, we are building intelligence on ground we cannot inspect.

As artificial intelligence (AI) models increasingly train on the output of other models, the lineage of data collapses into a fog. I argue that provenance, a signed and offline-verifiable chain of custody for synthetic data, is the only durable defence. This is the case for treating data lineage as infrastructure, not paperwork.

---

Originally published on mickai.co.uk. This is a cross-post; the canonical version, with the full body, footnotes and references, lives on the mickai.co.uk article page.