Your chats are not training data: a two-minute privacy check before you paste

Every AI company says the words "we take your privacy seriously", and at this point the phrase means nothing to anyone. So instead of saying it, here is the exact shape of what happens to your data at Dipole ML Inc., plus a two-minute checklist that works on us or on any other AI app.

What we do with your chats

The short version: nothing. Your conversations do not train our models in any tier, ever, and there is no settings screen where you have to keep it that way. The pipeline that would move your chats into training does not exist, which is a strange thing to be proud of, but here we are.

Inference happens in ephemeral memory. Once the response is back, what you sent is gone. The prompt, the completion, the code, the attachments. Saved conversations are the one exception, and even those sit under an individual Data Encryption Key that an engineer debugging the storage layer cannot see through. Usage metering sees how much you wrote, never what. The full architectural walkthrough, encryption and prompt-injection defenses included, lives in our earlier writeup on the Raven system.

The half nobody asks about: where the servers are

Retention is the part of privacy people ask about. Jurisdiction quietly decides more of it. Another country's legal processes can reach data sitting in that country, and they do not consult your contract first; provider terms also change the way terms always change, unilaterally, with a notice nobody reads. Our gateway and our serving infrastructure sit under Canadian jurisdiction, PIPEDA compliant and GDPR aligned, with no data broker wedged in the middle. "Where does the data live" is a one-sentence answer for us. We think every vendor should be able to manage a sentence.

The two-minute checklist for any AI app

Five questions, two minutes, and they work on us or on anyone else. A vendor that cannot answer them quickly, in writing, has told you something too.

• Who trains on what? At the big consumer labs, your chats have historically helped train the models unless you went digging through settings to say no. Opting out of a default was never supposed to be your job
• Is retention "zero" or "we keep it briefly"? Get the number. Then ask what happens to attachments and code, because that is where the interesting files live
• Who can see it? Staff, subcontractors, reviewers, "trusted partners": push on this one and the list keeps growing
• Where does it run? That one decides which laws and which legal processes can reach your data
• Can you delete it, backups included? A shrug on the backups half is its own answer

Our answers, for the record

For the record, ours: no training on your data in any tier, zero retention on inference, saved chats encrypted under per-user keys our staff cannot read, Canadian jurisdiction on infrastructure we operate ourselves, and deletion honored in full within a stated window. Each of those claims is written into our terms and privacy policy, kept in plain language on purpose so they can be checked rather than trusted.

Frontier intelligence should not cost you your privacy, and using it should not cost an afternoon in a settings menu. Ours costs nothing. That is the entire position.