Federated Learning Spreads the Training. It Does Not Spread the Trust.

Splitting a model across a thousand devices is a real privacy win. It does nothing to tell you whether the result can be trusted. For that you still need one tamper-evident anchor.

Federated learning distributes where training happens, and people quietly assume it distributes trust along with it. It does not. You can decentralise the computation completely and still be unable to prove what the system actually did. The fix is not more decentralisation. It is one signed, offline-verifiable record that no participant can rewrite.

---

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.