Blog · Essay
Household AI, and the data it should never train on
A household assistant sees your calendar, your money and your family's private arrangements. Privacy here is not a settings page, it is a business model: what the company earns from decides whether your data becomes training data. Sivella never trains on household content and never sells it, on any tier.
Think about what a household assistant would need in order to be useful. Not what it displays. What it holds.
It holds where each of you will be on Thursday. What you earn and what you owe. That one of you is job hunting quietly. Which relative you are not speaking to. What the doctor said. The address of the house and the hours it is empty. There is not much left over.
That is a more intimate corpus than most people's search history, and it arrives pre-sorted and labelled by the people it describes.
Ask what the company earns from
Every privacy question about an AI product reduces to one commercial question: where does the revenue come from. If it comes from you, your data is a cost to be protected. If it comes from someone else, your data is the product being sold, and every reassuring line in the policy is written to survive that fact.
Sivella charges a subscription and states plainly that this is deliberate: the business runs on subscriptions rather than on data. That is checkable against the pricing page in a way that a promise is not.
Two commitments follow from it, and both are worth reading in the policy rather than taking from an FAQ. Household content is never used to train, fine-tune or improve any AI model, on the free tier as well as the paid ones. And it is never sold or shared for advertising. Where an outside model provider is involved, that provider is under terms that forbid training on your content too, which matters because otherwise the promise stops at the first API call.
The free tier is where the seam usually is
Free is the tier that normally pays for itself some other way. A great many products draw the line here: paid customers get the strict handling, free users get a footnote saying their conversations may improve the service.
So check the free tier specifically, not the headline. Sivella's no-training commitment covers free and paid alike, which is the version of the claim that actually costs the company something.
Being able to look is the real control
A guarantee you cannot inspect is a guarantee you are taking on faith. The practical test of a household assistant is whether you can open what it thinks it knows about you and change it.
Sivella's memory is built to be inspected and edited by the household. You can see what it is holding, correct what is wrong, and remove what it should never have kept. That is different from a deletion request, and it is the control you will actually use, because most of what accumulates is not sinister, it is stale and slightly wrong.
Know how it ends before you begin
Read the exit before you commit real data. Sivella's is written down: asking to
delete deactivates the account immediately, a 120-day recovery window follows,
and after it your identifying information is irreversibly anonymized. A request
to [email protected] is handled within 30 days, and the same address
gets you a machine-readable archive of what is there.
Specific numbers and a named address are the tell. Vague reassurance costs nothing to write. A published window is a commitment somebody can be held to.