Private AI that never uploads your words — and how to check it yourself

Most privacy claims you read are true but not comparable. Here is the ladder, from weakest to strongest, so you can tell which one a product is actually offering.

• Policy privacy: your words go to a company's servers, and the company promises how it will treat them. You are trusting the policy, the people, and every future change to both.
• Account privacy: you can switch off training or memory, or delete history. Helpful, but your words still travel to the server when you ask, and the setting can move between plan tiers or be re-enabled by default in an update.
• Local privacy: the AI model runs on your own computer. Your question is read and answered on the machine in front of you. There is no server to send it to.

Only the third level means your words physically stay with you. That's the kind Konshus Sovereign provides for chat. The first two are worth having too — but they are promises, and a promise needs a company to keep it, indefinitely, across every ownership change and policy rewrite.

That sentence answers one narrow question: what happens to your words after they arrive. It doesn't answer the earlier question of whether they should arrive at all.

Think of it like mail. A company can promise to open your envelope, read it, and then burn it. That is a real promise, and it may even be honest. It still means your letter was read somewhere else. Local AI means the letter never left your desk.

There is also the matter of time. A policy applies today. Your reason for caring about privacy may last decades — a health question, a family situation, a business idea, a draft you never sent. The only guarantee that survives every future policy change is a model that never had a copy.

Plenty of people have questions they'd rather not type into a website: health worries, money, legal trouble, family situations, work that's under a confidentiality agreement, or simply thoughts that are nobody else's business.

Others care for practical reasons. Rural households, travelers and preparedness-minded families want an AI that keeps working when the connection doesn't — and they tend to value confidentiality more than most.

Some have a professional obligation. If you are a therapist, a lawyer, a clergy member, a doctor, or a manager handling sensitive people problems, the question isn't whether a provider is trustworthy in general. It's whether you can honestly tell a person that what they told you never left your machine. With a local brain, that sentence is true.

• Every question you ask and every answer you get in Sovereign's chat.
• The AI brain itself, downloaded once.
• Your private local Vault: the memories copied down from Konshus Online, stored encrypted on your computer, protected by a Recovery Key that only you hold.
• Anything you add to Sovereign locally.

Sovereign does not upload your chats automatically — not to Konshus Online, not anywhere. Copying runs one way: from your Online Vault down to your computer. If you want a Sovereign conversation in your Online Vault, you save it as a file yourself and add it on purpose.

Under each answer, "Why I know this" shows which entries in your Vault it drew on, so you can see the reasoning rather than guess at it.

Being honest about the exceptions matters more than a slogan. Sovereign uses the internet for:

• Downloading the app and the brain you choose.
• Linking to your Vault and copying memories down from Konshus Online when you ask it to resync.
• Checking whether a newer version exists. The app tells you and you choose whether to update; it never installs silently.

None of these send your chat. And you can block all of them with the Network switch. Every connection the app attempts is listed with the address it is trying to reach and a plain-language reason, so "we only connect for these things" is something you can audit rather than take on faith.

Sovereign also sends no analytics, no telemetry and no crash reports. There is nothing to switch off in a settings menu, because nothing is collected.

Nothing in this article asks you to trust a promise you cannot check. Every privacy claim below is something you can watch for yourself on your own computer.

• Turn off Wi-Fi or unplug your network cable. Ask Sovereign a question you have never asked. If an answer appears, it came from your computer.
• Use the Network switch in Sovereign to block the app's internet access, and confirm chat keeps working.
• Watch the connection list while you chat: if nothing is attempted, nothing left your machine.
• If you're comfortable with it, use your computer's own network monitor (Activity Monitor on Mac, Resource Monitor on Windows) while you chat.
• Ask it something about your life that only your Vault would know, with the internet off, and read which memories it says it used.

If a privacy claim can't be checked from your own computer, it isn't a privacy claim. It's a hope.

Local privacy moves the question from "can we trust the company?" to "is this machine safe?" That is a better position to be in, but it is not automatic.

Sovereign's local Vault is encrypted, and the 24-word Recovery Key that unlocks it is shown once and never stored by us. That means two things. First, a copied or stolen Vault file is useless without your key — including to us. Second, if you lose that key, there is no support ticket that gets the Vault back. Write it down and keep it with your other important papers.

Keep your computer's own login password strong and your disk encryption on. Local storage does not replace device security: it protects your words from servers, not from someone sitting at your keyboard.

Privacy includes being able to leave cleanly. Sovereign can erase its local Vault, conversations, brain and settings from inside the app. If the app won't open, a separate remover deletes the app, your private vault, the brain, settings, logs and its saved keys from your computer.

One honest limitation: your own backups. If you use Time Machine or another backup tool, copies made there are yours to manage — Sovereign doesn't reach into them. An old copy sitting in a backup can't be opened without the key that was deleted, but it is still your housekeeping, and we'd rather say so than quietly overpromise.

Local AI protects your words from servers. It cannot protect your computer from someone who has your password, or from malware already on it. Keep your computer updated, use a sign-in password, and store your Recovery Key somewhere safe and offline.

It also doesn't make the AI correct. A model answering privately can still be wrong — privately. For anything medical, legal, financial or safety-critical, verify with a real source. The privacy is real; the infallibility isn't.

And if someone else in your house uses the same computer, they can open the same app. That sounds obvious, and it's the most common privacy mistake in any household.

People imagine local privacy as a ritual: flipping switches, checking logs, being vigilant. In practice it's the opposite — it's the absence of a decision.

On a cloud service, privacy is something you configure and then forget to check. You turn off training in one menu, discover a memory setting in another, and wonder whether an update moved either of them. With a local brain, the question never comes up, because there's nothing to configure. You just ask.

The moments you do notice are the quiet ones. A friend asks which AI you use and you realize you don't have to explain a policy. You read about another company's breach and it doesn't touch you. You type something you'd have hesitated over elsewhere and nothing about the app changes.

That's the honest pitch. Not a dashboard full of toggles, and not a promise you have to re-read every year. Just the ordinary experience of a tool that doesn't send your words anywhere, so you stop thinking about it.

Partly, and it's worth knowing which parts. You give up live information — news, prices, weather — and you give up the very largest models, which still lead on the hardest reasoning. You also give up the fact that a cloud service works on any device you pick up.

What you keep is most of what people actually use AI for day to day: explaining, drafting, rewriting, summarizing, planning, thinking out loud, and asking about your own material. And you keep something cloud AI can't offer: an assistant that already knows you and doesn't need to be told again.

The practical answer for most people is both. Use a cloud AI when you need the live internet. Use a local one when the question is personal, confidential, or being asked somewhere without signal. With Konshus, the same Vault feeds both, so your AI knows you either way.