Your Key, Your Data: What Privacy-First AI Actually Means
“Privacy-first” has become the “all-natural” of software — a phrase that means whatever the person saying it wants it to mean. For an AI product, it usually decodes to “we send your data to a model provider, but we promise to be careful about it.”
That’s not nothing. But it’s a promise, not a property. When we built the privacy model for PhraseKit, we wanted something a user wouldn’t have to take on faith, because the thing we’re asking to see — their rough, unpolished writing before anyone else reads it — is genuinely personal.
Here’s how we actually think about it.
The three questions that matter
Strip away the marketing and privacy in an AI product comes down to three concrete questions:
- Where does my text go?
- Is it stored, and by whom?
- Is it used to train a model?
Most “privacy-first” claims answer only the third question, and only for the provider’s benefit. We tried to give an honest answer to all three, structurally.
Two doors, both private
PhraseKit runs in one of two modes, and you choose.
Managed inference. You sign in and we handle the model call. Your text goes to a model provider, gets a suggestion back, and that’s the end of its life — it isn’t persisted on PhraseKit servers. Crucially, we deliberately use provider tiers that do not train on customer inputs, and we keep an independent fallback provider so a single outage doesn’t take the app down or tempt us into a sloppier default. This is the easy on-ramp: nothing to configure, and still no data warehouse with your name on it.
Bring your own key. If you’d rather trust no one but your provider, you plug your own API key into the desktop app, and it talks directly to that provider. Your text never touches our infrastructure at all. There is no middleman to trust because there is no middleman.
Notice what’s the same in both: we are not building a corpus of your writing. The most private system is the one that never had your data to lose in the first place.
Why “no middleman” is an architecture, not a slogan
The temptation in a SaaS AI product is to route everything through your own backend so you can log, analyze, and “improve the experience.” Every one of those logs is a liability — a thing that can be breached, subpoenaed, or quietly repurposed later. The privacy-first move is to design so that the sensitive data simply isn’t there to collect.
That constraint costs you something. You give up some analytics. You have to solve abuse and rate-limiting without reading content. You have to make the subscription and quota system work off metadata instead of the text itself. We think that trade is worth it, and more importantly, we think it’s the trade users would make if they were in the room.
The takeaway
Privacy-first shouldn’t mean “trust us.” It should mean the system is built so there’s less to trust us with. If the honest answer to “where does my data go, is it stored, is it trained on” is “to a provider that won’t train on it, stored nowhere we control, and if you prefer, not through us at all” — that’s a property you can verify, not a promise you have to swallow.
We design this way because we build our own products and we’re our own first users. If you’re building an AI product and want the privacy story to be real instead of rhetorical, let’s talk.