Private by default: what it should actually mean

We started writing about private by default years before AI made the phrase fashionable. It has since become a badge on landing pages, which usually means it has stopped meaning anything. It is worth pinning down again, because for AI products the stakes are higher than they were for the apps that popularised the term.

The original idea

Ann Cavoukian's Privacy by Design made a claim that sounded radical in the 2000s and now sounds obvious: privacy and value are not a trade-off. The trade-off framing was invented by businesses whose model required the trade. Designed properly, a system can give people more value precisely because they trust it with more.

Private by default is the operational half of that idea. It says the safe state is the starting state. Nothing is shared unless a person does something deliberate to share it.

What it means for an AI product

The default state of most AI products is the opposite. You upload, and the product's terms establish broad rights over what you uploaded. Privacy is the thing you configure afterwards, if you find the setting.

A product that actually starts private looks like this:

• Nothing you upload trains a shared model. Not by default, not by opt out, not at all.
• Nothing is publicly reachable until you publish it, deliberately, with a link you created.
• Sensitive material can be invite only, so specific people get access and nobody else does.
• Access you granted can be taken back, and taking it back actually works.
• Deletion removes the thing, and the vendor can tell you exactly what removal covers.

The part people skip: honesty about the edges

The reason private by default became a slogan is that the honest version has edges, and edges do not fit on a landing page.

Here is one of ours. Regional processing preference is real, and for some models there is no regional path at all, which means those queries fall back to US processing. We could avoid saying that. Saying it is the difference between a commitment and a slogan, because a commitment tells you where it stops.

Any vendor whose privacy story has no edges has either not built the product yet or is not describing the one they built.

Public when you choose it

Private by default is not private forever. Most people building an AI Twin want it to be found. The point is the order of operations: your knowledge is private, and then you decide what becomes reachable, to whom, and when.

An AI Twin with a public link is a deliberate act of publishing. An AI Twin holding your client files is not. Both live in the same product, and the product should never confuse the two.

Why it decides whether people build at all

The material worth putting into an AI Twin is exactly the material people are most careful with. Nobody hesitates over uploading a blog post. They hesitate over the case studies, the frameworks, the unpublished work that took a decade.

If the default is unsafe, that material never gets uploaded, and the product is only ever useful for things that did not need protecting. Private by default is not a feature bolted onto a knowledge product. It is the precondition for anyone using it seriously.