Private AI and data control

What "private AI" can actually mean

The phrase covers four different properties that get bundled together in marketing and come apart in practice. A product can have any of them without the others.

Not used for training

Your conversations do not become training data for a future model. This is the one most people mean, and it is a policy commitment rather than a technical guarantee — you are trusting the operator.

Not retained

Content is deleted rather than stored. Genuine zero retention conflicts with conversation history, so most products that offer it do so as a mode you turn on and lose features for.

Not linked to a legal identity

Whether the account can be traced to a named person, which in practice is decided by how you pay far more than by anything else.

Local

The model runs on hardware you control and nothing leaves your machine. This is the only version that is a technical guarantee rather than a promise, and it costs a GPU.

A hosted service can offer the first three honestly. It cannot offer the fourth, and any hosted product claiming it is describing something else.

What Tartarus AI stores

Stated plainly, because a vague answer here is itself an answer.

What crypto billing does and does not buy

Paying in cryptocurrency means there is no card number, billing address or stored payment profile associated with your account. That is a real reduction in what exists to be leaked, subpoenaed or sold, and it is why the product is built this way rather than as a marketing flourish.

It is not anonymity. Most chains are public ledgers, and a payment made from an exchange account that performed identity verification is linkable by anyone who can see both sides. Monero behaves differently from Bitcoin here. If unlinkability is your actual requirement, the choice of currency matters more than the choice of platform.

It also does not affect anything else on this page. The email address, the conversations and the server logs exist regardless of how you paid.

Where your messages actually go

A message you send is passed to the inference service that produces the reply. That service runs on hardware we operate rather than a resold third-party endpoint, which is why the model configuration is ours to set and why conversations do not leave the boundary to be used by someone else.

Two exceptions are worth knowing. When a request uses web search, the search terms the model chooses go to the search provider and the pages it selects are fetched — that traffic leaves, though nothing identifying you accompanies it. And if you sign in with Google, that sign-in is between you and Google in the usual way.

Which threat model this fits

Being specific about who this protects you from is more useful than a reassurance.

Reasonable fit

You do not want your conversations training a model, do not want a card number and billing address held against your account, do not want your work read by an automated moderation system, and want to be able to delete things and have them gone.

Poor fit

You need a guarantee that no operator can ever read your content, you are defending against a state-level adversary, or you are handling regulated data that requires a processing agreement. Run a model locally, or use a vendor that will sign the paperwork.

The honest summary: this is a hosted service, so the operator is in your trust boundary. What we can do is hold as little as the product allows and say exactly what that is.

What you control

The practical levers available to you from inside the account.

• Delete any conversation, which removes it from your account immediately and purges it from storage thereafter.
• Delete the account, which removes your conversations and messages. Payment records are kept where accounting rules require it.
• Sign up with an email address that is not tied to your name — nothing in the product requires a real one.
• Choose the currency you pay in, which is the decision that most affects linkability.
• Revoke any API token immediately, without affecting the others.