Answers · a straight answer, sourced
Is running AI locally actually private?
Local inference is necessary for privacy but not sufficient: the model runs on your machine, but the application around it decides whether your prompts, memory, or documents ever leave — so "runs locally" and "is private" are two separate claims you have to check separately.
A local model means the tokens are generated on your hardware. That rules out the biggest leak, but it says nothing about the rest of the app: telemetry, update checks, cloud "sync", crash reporters, or a settings screen that quietly falls back to a hosted model. A product can run a model locally and still phone home constantly.
The test that actually separates the two is disclosure at the moment it happens. Ask: when this app sends something off my machine, does it tell me then — on screen, at the point of use — or only in a policy PDF? A product that shows you the exact bytes before a remote call is proving the claim; one that asks you to trust a toggle labelled "improve the product" is not. This is the third of the three questions worth asking any private-AI product.
It is also why local is not the finish line. The honest version of the claim is: local by default, remote only when you choose, and shown to you the instant anything crosses your boundary. SovereignAI does this with a customs declaration — before any message goes to a remote provider, the precise outgoing context is shown for approval, and every remote answer carries a receipt of what left. Local endpoints never ask, because nothing leaves.
The parts that are not private, we name in the Sovereignty Ledger: the database is not encrypted at rest today (the honest mitigation is full-disk encryption), and even open-weight models are a third party’s artifact. "Private" here means custody and disclosure, not a magic word.
Own the answer, not just read it
SovereignAI is private AI you run yourself: local models, memory with receipts, disclosure at the moment anything leaves, and an exit you can verify. The trial is one Docker command, and deleting the volume deletes every trace of you.