The Room
A room isn’t a door. A room is where you live.
There is a sentence that lives in the middle of the Private Assistant product page: “Your conversations stay on your device.”
Most people read it as a reassurance. A lock on the door. We won’t take what’s inside.
That reading isn’t wrong. But it’s incomplete.
A door is about the threshold. Who gets in, who stays out. The whole conversation about privacy in AI has been a conversation about doors — who has the key, whether the walls are thick enough, what happens if someone breaks through. These are real questions. They matter. We’ve thought hard about them and answered them honestly: the model runs on your device, the conversation never leaves, there is no server in the middle watching you think.
But a room isn’t a door. A room is where you live.
Here is something worth knowing about how people think: they don’t think in their heads — not only there, anyway. They think in sentences. They think out loud. They think in the act of telling someone else — a friend, a therapist, a notebook — and the telling is part of the thinking. You don’t always know what you believe until you’ve tried to explain it.
This is the part of conversation that gets lost in the anxiety about data.
When you have a conversation with an AI that lives on someone else’s server, something invisible happens. The conversation is audited as it goes. Not necessarily read — maybe not even stored — but it passes through infrastructure that belongs to someone with different interests than yours. And you know this, at some level, even if you never think about it explicitly. You calibrate. You self-edit. You skip the half-formed questions. You don’t ask the ones that might be embarrassing, or that might reveal something you’re not sure you believe yet.
Is this job actually making me unhappy, or am I just tired?
What do I think about this decision I’ve already made?
Do I actually want this, or have I just convinced myself I do?
This is not paranoia. It’s just how we work in observed spaces. We’ve always done it.
What you lose, when you calibrate like this, is not safety. You lose access to your own unfinished thinking.
Private Assistant doesn’t fix this by building better walls. It fixes it by putting the room where it belongs — on your device, in your hands.
When the model runs on your device, there is no threshold to protect. The conversation isn’t transmitted, so there’s nothing to intercept. You don’t have to trust our server administration — we don’t have one. You’re not extending a privilege to us. You’re not granting access. You’re just thinking.
Which means you can be tentative. Uncertain. Contradictory. You can say I don’t know if this is even a real concern and find out if it is. You can be wrong about what you’re trying to say and talk your way toward something better. You can change your mind in the middle of a sentence and not have that change be part of a dataset that says something about you.
The autonomy we’re protecting isn’t just about your data. It’s about your thinking.
We spent a long time making the architecture argument: the model is good, the system is secure, the privacy guarantee is technical rather than policy-based. That argument is true and we stand behind it.
But architecture tells you about the walls. It doesn’t tell you what it feels like to live inside them.
The Room is what it feels like. You show up with half an idea and a nagging feeling and a question you couldn’t quite phrase, and you think out loud until you know what you actually believe. And then you close the app and you go do something with what you found.
It was yours. The whole time, it was yours.
Your conversations stay on your device.
We meant that literally. We also meant everything else.
Part II of “Architecture & Autonomy.” Part I: Trust as Architecture.
Your AI. Your device. Your room.
Explore Our Products