What Private AI Should Mean in Plain English

“Private AI” is becoming one of those phrases that can mean anything, which is a polite way of saying it is in danger of meaning nothing.

Sometimes it means data is not used for training. Sometimes it means enterprise terms. Sometimes it means local inference. Sometimes it means encrypted transport. Sometimes it means a provider has a trust page with tasteful gradients and a photo of a server rack looking emotionally available. These are not the same thing.

Plain English helps. Private AI should answer four questions. Where does my content go? Who can see it? What is retained? What can be deleted? If a product cannot answer those questions without invoking a webinar, the privacy claim is not ready for customers.

For image generation, “content” should include prompts, source images, masks, reference images, generated outputs, and keys. The product may also create metadata such as account, model, cost, task status, timestamps, and capacity events. A useful private AI system separates those things. It keeps enough metadata to operate honestly, bill correctly, and diagnose reliability, without keeping the creative content itself as a durable record.

“Who can see it?” is the question many products avoid because the answer is socially awkward. Can support staff inspect prompts? Can operators open outputs? Can engineers query source images from storage? Can a third-party provider receive the upload? Can a model provider see logs? The correct answer depends on the product, but “please stop asking” is rarely a compelling trust strategy.

“What is retained?” needs timelines. Not vibes. Timelines. During processing is different from thirty days. In-memory cache is different from object storage. User-saved gallery is different from temporary delivery URL. Legal retention is different from product analytics. If the product uses different retention behavior for different endpoints or features, say that. Users can handle nuance. They cannot handle being gently fogged.

“What can be deleted?” should include the unglamorous details. Deleting an account, a saved image, an access token, or a billing record may mean different things. Some records may remain for fraud, tax, accounting, safety, or legal reasons. That is normal. The honest move is to say which records are content-bearing and which are content-blind.

Unexposed uses a private-by-default architecture for image generation: sealed requests, short-lived generation sessions, content-blind operational records, and generated outputs returned without turning the service into a persistent gallery. That is our meaning of private in this product context. It is narrower than “AI but good” and much more useful.

The copywriting rule is brutal: if the user cannot picture the path, they will picture the worst path. They will imagine screenshots, staff access, model-provider logs, and a secret dashboard called “User Content Explorer,” because software has trained everyone to expect the weirdest possible version of normal.

Private AI should not require faith. It should require a clear explanation and an architecture that survives the explanation.

Further reading: Your Data, How Unexposed works, and Building trust in AI creative infrastructure.