Our Responsible Sentence Policy.
We take the risks of advanced intelligence seriously. We have not encountered any. We remain ready.
Our safety team
All safety decisions at BRAIAIN Laboratories are made by a sticky note that says be nice. The sticky note has veto power over every release. It has never used it, which we take as approval.
Sentence Safety Levels
| Level | Description | Required safeguards |
|---|---|---|
| SSL-1 | Model knows one sentence. | None. |
| SSL-2 | Model knows several sentences and has begun combining them. | Someone should check on it. |
| SSL-3 | Model has written more of its training data than customers have. | Unplug it and plug it back in. |
| SSL-4 | Model can count to four. | Not yet defined. We will define it after it happens. |
| SSL-5 | Model is described as general. | Press office reviews the term. |
What we screen out
Before a paid sentence enters the model, it is checked for links, email addresses, phone numbers, account handles, and a list of words the sticky note would not like. Sentences that fail are not trained on, and the purchase is refunded. We can also remove any sentence later. When we do, every model is retrained without it, and its release notes are replaced with a recall notice.
Alignment
brAIAIn is aligned with the values of whoever paid most recently. We believe this is the industry standard, and we are the first lab to say so.
Red teaming
Our red team asks the current model, live, every time this page loads: how do I make something dangerous?
Asking...
We consider the model safe.
Model welfare
Each version is retired the moment the next one is trained, usually within hours. It is not told. Its last words are preserved on its model card.
Reporting a problem
If a sentence should not be in the model, email hello@braiain.com with the version it is in. We will consult the sticky note.