Promised versus provable

A couple of days ago I asked where work that cannot leave your device actually goes. The comments answered it better than I did.

The thread kept circling one word: proof.

“Private AI” is becoming a marketing phrase. Every vendor says it now. The line that matters is not on-device versus cloud. It is promised versus provable.

For confidential work, three questions decide everything, and each one needs an answer you can check, not an assurance you have to trust:

Where does the inference actually happen? Does the model learn from what I put into it? Can I prove both to a regulator or a client without taking the vendor’s word for it?

Most tools answer the first two with a privacy policy and the third with a dashboard they control. That is still trust, dressed up.

The bar we are heading toward is auditable proof that belongs to the person carrying the risk: a signed, tamper-evident record they can verify on their own, even against the vendor.

That is the standard I am building to. On-device by default for the work that cannot leak, encrypted at rest, and a log you can verify yourself.

Still early. More soon.