Technology
Generative AI is a production tool for us, not a slogan. This page is a plain description of what runs where, what a human signs off on, and what we refuse to do with it.
The pipeline
Authored constraints first
Every generative system we run reads from a specification a human wrote: the art bible for a game, the domain model for an application, the tone guide for both. Models are asked to fill a shaped hole, never to invent the shape.
Runtime direction
A planner tracks state and decides what happens next — which authored beat a game schedules, which step an application takes — and what the generated content around it is allowed to be.
Human review before release
Generated art, audio, copy and code that ships in a release passes a review queue. Content generated live during use runs behind safety filters and stays inside the authored constraint set.
Provenance and training data
We license or own the material used to tune our style models, and we keep records of it. We do not train on customer conversations, documents or account data.
What we will not do
- We do not train on customer data, chat logs, documents or save files.
- We do not generate likenesses or voices of real people without a signed release.
- We do not ship generated content into a release without a human reviewing it.
- We do not use generative systems to set prices, target spending, or personalise monetisation.
- We do not use automated processing to make decisions about a person with legal or similarly significant effects.